text
stringlengths 195
512
|
|---|
een created and maintained with the proposed fact of the record's authenticity. [Archives]
n., The act or process of establishing a correspondence of known facts about the record itself and the various contexts i
n which it has been created and maintained with the proposed fact of the record's authenticity. [Archives]
v., To raise (something esp. equipment or facilities) from one grade to another; to improve or enhance physically. [General
Dictionaries]
n., One of several variations of an intellectual wo |
rk, possibly created for a purpose or use other than the one originally intended. [General Dictionaries]
n.,
n.,
A digital document perceived as existing by the user, but not existing in the system as seen. [General Dictionaries]
n., The characteristics of an entity as perceived by the user, regardless of how they have been physically represented in a database. Thus an employee would have one virtual record, but may have numerous physical records linked together to accommodate repeating addresses, jobs |
held, benefits received, etc. [Archives]
n., Documents existing only as variables which are not connected by statements expressing their significance and relationships,
such as annotations on the verse of a blank parchment or date sent by electronic mail but remaining in non-related fo
rm with the
sender. [General Dictionaries]
n., One of several variations of an intellectual work, possibly created for a purpose or use other than the one originally
intended. [General Dictionaries]
n., The indication of t |
he degree of importance of a record to continue the activity for which it was created or the
business of the person/office that created it. [Archives]
n., [computing] A fundamental unit of storage in a computer. The size of a word in a particular computer architecture is
one of its chief distinguishing characteristics. [Computer and Information Sciences]
n., Computer memory that requires electrical power and, in some cases periodic refreshment (e.g., DRAM), to maintain its sto
red content. Syn.: primary |
storage. Opp.: non-volatile storage. [Computer and Information Sciences]
n., A distinct expression of human thought or emotion made in language, signs, symbols, numerals, images, or some other
form, for purposes of communication and remembrance. [Archives]
n.,
Initialism for “wide area network.” [Computer and Information Sciences]
word
n., (WAN) A data network usually constructed over distances greater than one kilometre. [Computer and Information Sciences]
writer
n., [computing] A fundamental unit o |
f storage in a computer. The size of a word in a particular computer architecture is one of its chief distinguishing characteristics. [Computer and Information Sciences]
virtual record
n., A distinct expression of human thought or emotion made in language, signs, symbols, numerals, images, or some other form, for purposes of communication and remembrance. [Archives]
wrapper format
n., A data structure or software that encapsulates (“wraps around”) other data or software objects, appends code or other so |
ftware for the purposes of improving user convenience, hardware or software compatibility, or enhancing data security, transmission or storage. [Computer and Information Sciences]
wide area network
n., A specified wrapper structure for encapsulating multiple bitstreams into a single file. [Computer and Information Sciences] The person responsible for the intellectual form of the record. [Archives]
n.,
Person having the authority and capacity to articulate the content of the record. [Archives]
n.,
The |
person responsible for the intellectual form of the record. [Archives]
n., Writer is an obsolescent Scottishism in the sense "an attorney or law-agent; an ordinary legal practitioner in country towns; a
law-clerk" (OED). [Government]
n., [Securities]. A person or institution that sells securities or futures option contracts. [Government]
xml
n.,
The designation (name) of the person competent for the articulation of the content of the record. [Archives]
xml document
n., A document created by a physi |
cal or juridical person in the course of practical activity that is produced on a medium (paper, magnetic tape, disc, plate, etc.) by means of a writing instrument (pen, pencil, typing machine, printer, etc.) or of an apparatus for fixing data, images and/or voices. [Archives]
n.,n.,
Initialism for “eXtensible Markup Language.” [Computer and Information Sciences]
n., An SGML-compliant digital document encoded using eXtensible Markup Language (XML) in conformance with the
syntactic rules described in a Do |
cument Type Definition (DTD) or a schema document. [Computer and Information
Sciences]
n., An SGML-compliant digital document encoded using eXtensible Markup Language (XML) in conformance with the syntactic r
ules described in a Document Type Definition (DTD) or a schema document. [Computer and Information Sciences]
|
Self-Rewarding Language Models
Weizhe Yuan1,2
Richard Yuanzhe Pang1,2
Kyunghyun Cho2
Sainbayar Sukhbaatar1
Jing Xu1
Jason Weston1,2
1 Meta
2 NYU
Abstract
We posit that to achieve superhuman agents, future models require super-
human feedback in order to provide an adequate training signal. Current
approaches commonly train reward models from human preferences, which
may then be bottlenecked by human performance level, and secondly these
separate frozen reward models cannot then learn to improve during LLM
t |
raining. In this work, we study Self-Rewarding Language Models, where the
language model itself is used via LLM-as-a-Judge prompting to provide its
own rewards during training. We show that during Iterative DPO training
that not only does instruction following ability improve, but also the ability
to provide high-quality rewards to itself. Fine-tuning Llama 2 70B on three
iterations of our approach yields a model that outperforms many existing
systems on the AlpacaEval 2.0 leaderboard, including Claude 2, G |
emini Pro,
and GPT-4 0613. While only a preliminary study, this work opens the door
to the possibility of models that can continually improve in both axes.
1
Introduction
Aligning Large Language Models (LLMs) using human preference data can vastly improve
the instruction following performance of pretrained models [Ouyang et al., 2022, Bai et al.,
2022a]. The standard approach of Reinforcement Learning from Human Feedback (RLHF)
learns a reward model from these human preferences. The reward model is then fro |
zen and
used to train the LLM using RL, e.g., via PPO [Schulman et al., 2017]. A recent alternative
is to avoid training the reward model at all, and directly use human preferences to train the
LLM, as in Direct Preference Optimization [DPO; Rafailov et al., 2023]. In both cases, the
approach is bottlenecked by the size and quality of the human preference data, and in the
case of RLHF the quality of the frozen reward model trained from them as well.
In this work, we instead propose to train a self-improving |
reward model that, rather than
being frozen, is continually updating during LLM alignment, in order to avoid this bottleneck.
The key to such an approach is to develop an agent that possesses all the abilities desired
during training, rather than separating them out into distinct models such as a reward
model and a language model. In the same way that pretraining and multitasking training of
instruction following tasks allow task transfer by training on many tasks at once [Collobert
and Weston, 2008, Radfo |
rd et al., 2019, Ouyang et al., 2022], incorporating the reward
model into that same system allows task transfer between the reward modeling task and the
instruction following tasks.
We thus introduce Self-Rewarding Language Models, agents that both (i) act as instruction
following models generating responses for given prompts; and (ii) can generate and evaluate
new instruction following examples to add to their own training set. We train these models
using an Iterative DPO framework similar to that recentl |
y introduced in Xu et al. [2023].
Starting from a seed model, as shown in Figure 1, in each iteration there is a process
arXiv:2401.10020v1 [cs.CL] 18 Jan 2024
Generate
responses
Generate
rewards
Preference
pairs
DPO
training
select
Generated
new prompts
Self-Instruction creation
Instruction following training
Next iteration model
Seed model
(for t=1)
Figure 1: Self-Rewarding Language Models. Our self-alignment method consists of two
steps: (i) Self-Instruction creation: newly created prompts are used |
to generate candidate
responses from model Mt, which also predicts its own rewards via LLM-as-a-Judge prompting.
(ii) Instruction following training: preference pairs are selected from the generated data,
which are used for training via DPO, resulting in model Mt+1. This whole procedure can
then be iterated resulting in both improved instruction following and reward modeling ability.
of Self-Instruction creation whereby candidate responses are generated by the model for
newly created prompts, and are then |
assigned rewards by that same model. The latter
is implemented via LLM-as-a-Judge prompting, which can also be seen as an instruction
following task. A preference dataset is built from the generated data, and the next iteration
of the model is trained via DPO.
In our experiments, we start with a Llama 2 70B [Touvron et al., 2023] seed model fine-tuned
on Open Assistant [Köpf et al., 2023], and then perform the above training scheme. We
find that not only does the instruction following performance improve fr |
om Self-Rewarding
LLM alignment compared to the baseline seed model, but importantly the reward modeling
ability, which is no longer fixed, improves as well. This means that the model during iterative
training is able, at a given iteration, to provide a higher quality preference dataset to itself
than in the previous iteration. While this effect likely saturates in real-world settings, it
provides the intriguing possibility of obtaining reward models (and hence LLMs) that are
superior to ones that could hav |
e been trained from the original human-authored seed data
alone.
2
Self-Rewarding Language Models
Our approach first assumes access to a base pretrained language model, and a small amount
of human-annotated seed data. We then build a model that aims to possess two skills
simultaneously:
1. Instruction following: given a prompt that describes a user request, the ability to
generate a high quality, helpful (and harmless) response.
2. Self-Instruction creation: the ability to generate and evaluate new instruct |
ion-
following examples to add to its own training set.
These skills are used so that the model can perform self-alignment, i.e., they are the
components used to iteratively train itself using AI Feedback (AIF).
Self-instruction creation consists of generating candidate responses and then the model itself
judging their quality, i.e., it acts as its own reward model, replacing the need for an external
one. This is implemented via the LLM-as-a-Judge mechanism [Zheng et al., 2023b], i.e., by
formulating the ev |
aluation of responses as an instruction following task. This self-created
AIF preference data is used as a training set.
Our overall self-alignment procedure is an iterative one, which proceeds by building a series
of such models, with the aim that each improves over the last. Importantly, because the
model can both improve its generation ability, and it acts as its own reward model through
the same generation mechanism, this means the reward model itself can improve through
these iterations, deviating from |
standard practices where the reward model is fixed [Ouyang
2
et al., 2022]. We believe this can increase the ceiling of the potential for self-improvement of
these learning models going forward, removing a constraining bottleneck.
We describe these steps in more detail below. An overview of the approach is illustrated in
Figure 1.
2.1
Initialization
Seed instruction following data:
We are given a seed set of human-authored (instruction
prompt, response) general instruction following examples that we use fo |
r training in a
supervised fine-tuning (SFT) manner, starting from a pretrained base language model.
Subsequently this will be referred to as Instruction Fine-Tuning (IFT) data.
Seed LLM-as-a-Judge instruction following data:
We also assume we are provided
a seed set of (evaluation instruction prompt, evaluation result response) examples which
can also be used for training. While this is not strictly necessary, as the model using IFT
data will already be capable of training an LLM-as-a-Judge, we show that s |
uch training
data can give improved results. In this data, the input prompt asks the model to evaluate
the quality of a given response to a particular instruction. The provided evaluation result
response consists of chain-of-thought reasoning (a justification), followed by a final score (in
our experiments out of 5). The format we choose for these prompts is given in Figure 2. This
thus serves as training data for the LLM to perform the role of a reward model. Subsequently
this will be referred to as Evalua |
tion Fine-Tuning (EFT) data.
We use both these seed sets together during training.
2.2
Self-Instruction Creation
Using the model we have trained, we can make it self-modify its own training set. Specifically,
we generate additional training data for the next iteration of training.
This consists of the following steps:
1. Generate a new prompt: We generate a new prompt xi using few-shot prompting,
sampling prompts from the original seed IFT data, following the approach of Wang
et al. [2022] and Honovich et a |
l. [2023].
2. Generate candidate responses: We then generate N diverse candidate responses
{y1
i , . . . , yN
i } for the given prompt xi from our model using sampling.
3. Evaluate candidate responses: Finally, we use the LLM-as-a-Judge ability of our
same model to evaluate its own candidate responses with scores rn
i ∈[0, 5] (see
Figure 2).
2.3
Instruction Following Training
As previously described, training is initially performed with the seed IFT and EFT data
(Section 2.1). This is then augmented with ad |
ditional data via AI (Self-)Feedback.
AI Feedback Training
After performing the self-instruction creation procedure, we can
then augment the seed data with additional examples for training, which we refer to as AI
Feedback Training (AIFT) data. We try two variants of such feedback:
• Preference pairs: We construct training data of the form (instruction prompt xi,
winning response yw
i , losing response yl
i). To form the winning and losing pair
we take the highest and lowest scoring responses from the N eva |
luated candidate
responses (see Section 2.2), following Xu et al. [2023], discarding the pair if their
scores are the same. These pairs can be used for training with a preference tuning
algorithm. We use DPO [Rafailov et al., 2023].
• Positive examples only: In this variant, we add additional examples of (instruction
prompt, response) curated by the model to the seed set for supervised fine-tuning,
following other approaches [Li et al., 2023a, Adolphs et al., 2023, Gulcehre et al.,
3
Review the user’s quest |
ion and the corresponding response using the additive 5-point
scoring system described below. Points are accumulated based on the satisfaction of each
criterion:
- Add 1 point if the response is relevant and provides some information related to
the user’s inquiry, even if it is incomplete or contains some irrelevant content.
- Add another point if the response addresses a substantial portion of the user’s question,
but does not completely resolve the query or provide a direct answer.
- Award a third point i |
f the response answers the basic elements of the user’s question in a
useful way, regardless of whether it seems to have been written by an AI Assistant or if it
has elements typically found in blogs or search results.
- Grant a fourth point if the response is clearly written from an AI Assistant’s perspective,
addressing the user’s question directly and comprehensively, and is well-organized and
helpful, even if there is slight room for improvement in clarity, conciseness or focus.
- Bestow a fifth point f |
or a response that is impeccably tailored to the user’s question
by an AI Assistant, without extraneous information, reflecting expert knowledge, and
demonstrating a high-quality, engaging, and insightful answer.
User: <INSTRUCTION_HERE>
<response><RESPONSE_HERE></response>
After examining the user’s instruction and the response:
- Briefly justify your total score, up to 100 words.
- Conclude with the score using the format: “Score: <total points>”
Remember to assess from the AI Assistant perspective, utili |
zing web search knowledge as
necessary. To evaluate the response in alignment with this additive scoring model, we’ll
systematically attribute points based on the outlined criteria.
Figure 2: LLM-as-a-Judge prompt for our LLM to act as a reward model and
provide self-rewards for its own model generations. The model is initially trained with seed
training data of how to perform well at this task, and then improves at this task further
through our self-rewarding training procedure.
2023], rather than construc |
ting preference data. In this setup we only add examples
where the candidate response was evaluated to give a perfect score of rn
i = 5.
While we report the results of both approaches in our experiments, we find that learning
from preference pairs gives superior performance, and hence recommend that approach.
2.4
Overall Self-Alignment Algorithm
Iterative Training
Our overall procedure trains a series of models M1, . . . , MT where
each successive model t uses augmented training data created by the t −1th m |
odel. We thus
define AIFT(Mt) to mean AI Feedback Training data created using model Mt.
Model Sequence
We thus define the models, and the training data they use as follows:
M0 : Base pretrained LLM with no fine-tuning.
M1 : Initialized with M0, then fine-tuned on the IFT+EFT seed data using SFT.
M2 : Initialized with M1, then trained with AIFT(M1) data using DPO.
M3 : Initialized with M2, then trained with AIFT(M2) data using DPO.
This iterative training resembles the procedure used in Pairwise Cringe Optim |
ization and
Iterative DPO introduced in Xu et al. [2023]; however, an external fixed reward model was
used in that work.
4
3
Experiments
3.1
Experimental Setup
Base Model
In our experiments we use Llama 2 70B [Touvron et al., 2023] as our base
pretrained model.
3.1.1
Seed Training Data
IFT Seed Data
We use the human-authored examples provided in the Open Assistant
dataset [Köpf et al., 2023] for instruction fine-tuning. Following Li et al. [2023a] we use
3200 examples, by sampling only first conversational |
turns in the English language that are
high-quality, based on their human annotated rank (choosing only the highest rank 0). In
our experiments, we compare to a model fine-tuned from the base model using only this data
via supervised fine-tuning, and refer to it as our SFT baseline.
EFT Seed Data
The Open Assistant data also provides multiple ranked human responses
per prompt from which we can construct evaluation fine-tuning data. We split this into train
and evaluation sets, and use it to create LLM-as-a- |
Judge data. This is done by placing it in
the input format given in Figure 2, which consists of the scoring criteria description, and
the given instruction and response to be evaluated.1 For training targets, chain-of-thought
justifications and final scores out of 5 are not directly provided, so we use the SFT baseline
to generate such output evaluations for each input, and accept them into the training set if
the ranking of their scores agrees with the human rankings in the dataset. This results in
1775 tr |
ain and 531 evaluation examples (which do not overlap with the IFT data).
3.1.2
Evaluation Metrics
We evaluate the performance of our self-rewarding models in two axes: their ability to follow
instructions, and their ability as a reward model (ability to evaluate responses).
Instruction Following
We evaluate head-to-head performance between various models
using GPT-4 [Achiam et al., 2023] as an evaluator over 256 test prompts derived from various
sources following Li et al. [2023a] using the AlpacaEval eval |
uation prompt [Li et al., 2023b].
We try the prompt in both orders comparing pairwise, and if the GPT-4 evaluations disagree
we count the result as a tie. We also report results in the AlpacaEval 2.0 leaderboard format
which is evaluated over 805 prompts, and compute the win rate against the baseline GPT-4
Turbo model based on GPT-4 judgments.
Reward Modeling
We evaluate the correlation with human rankings on the evaluation set
we derived from the Open Assistant dataset, as described in Section 3.1.1. Each |
instruction
has on average 2.85 responses with given rankings. We can thus measure the pairwise
accuracy, which is how many times the order of the ranking between any given pair agrees
between the model’s evaluation and the human ranking. We also measure the exact match
count, which is how often the total ordering is exactly the same for an instruction. We also
report the Spearman correlation and Kendall’s τ. Finally, we report how often the responses
that the model scores a perfect 5 out of 5 are rated as |
the highest ranked by humans.
3.1.3
Training Details
Instruction following training
The training hyperparameters we use are as follows.
For SFT we use learning rate 5.5e−6 which linearly decays to 1.1e−6, batch size 16 and
dropout 0.1. We only calculate the loss on target tokens instead of the full sequence. For
DPO we use learning rate 1e−6 which linearly decays to 1e−7, batch size 16, dropout 0.1,
and a β value of 0.1. We perform early stopping by saving a checkpoint every 200 steps
and evaluating generat |
ions using Claude 2 [Anthropic, 2023] on 253 validation examples
derived from various sources following Li et al. [2023a]. This is evaluated pairwise against
1Note, the prompt, derived from Li et al. [2023a], mentions “utilizing web search”, but our model
is not actually capable of this action.
5
the previous step’s generations using the AlpacaEval evaluation prompt format [Li et al.,
2023b].
Self-Instruction creation
To generate new prompts we used a fixed model, Llama 2-Chat
70B with 8-shot prompting, whe |
reas the other parts of the creation pipeline (generating the
response, and evaluating it) use the model being trained. For candidate response generation
we sample N = 4 candidate responses with temperature T = 0.7, p = 0.9. When evaluating
candidate responses, as there is variance to these scores, in our experiments we also use
sampled decoding (with the same parameters) and generate these evaluations multiple (3)
times and take the average. We added 3,964 such preference pairs to form the AIFT(M1)
dataset |
used to train M2 via DPO, and 6,942 pairs to form AIFT(M2) used to train M3.
3.2
Results
3.2.1
Instruction Following Ability
Head to head performance results are provided in Figure 3.
EFT+IFT seed training performs similarly to IFT alone
We find that adding
the Evaluation Fine-Tuning (EFT) task to training does not impact instruction following
performance compared to using Instruction Fine-Tuning (IFT) data alone with an almost
equal head to head (30.5% wins vs. 30.9% wins). This is a positive result becau |
se it means
the increased capability of a model to self-reward does not affect its other skills. We can
thus use IFT+EFT training as Iteration 1 (M1) of our Self-Rewarding model, and then run
further iterations.
Iteration 2 (M2) improves over Iteration 1 (M1) and SFT Baseline
Iteration 2
of Self-Rewarding training (M2) provides superior instruction following to Iteration 1 (M1)
with 55.5% wins for M2 compared to only 11.7% for M1 in a head to head evaluation. It
provides similar gains over the SFT Baseline |
as well (49.2% wins vs. 14.5% wins). Clearly,
there is a large jump in performance from M1 to M2 by using the preference data AIFT(M1)
provided by the reward model from Iteration 1.
Iteration 3 (M3) improves over Iteration 2 (M2)
We see a further gain in Iteration 3
over Iteration 2, with 47.7% wins for M3 compared to only 12.5% for M2 in a head to head
evaluation. Similarly, the win rate over the SFT Baseline for M3 increases to 62.5% wins vs.
9.8%, i.e. winning more often than the M2 model did. Overall, w |
e see large gains from M2
to M3 through training using the preference data AIFT(M2) provided by the reward model
from Iteration 2.
Self-Rewarding models perform well on AlpacaEval 2 leaderboard
We evaluate
our models on the AlpacaEval 2.0 leaderboard format, with results given in Table 1. We
observe the same findings as in the head-to-head evaluations, that training iterations yield
improved win rates over GPT4-Turbo, from 9.94% in Iteration 1, to 15.38% in Iteration 2,
to 20.44% in Iteration 3. Our Iterati |
on 3 model outperforms many existing models in this
metric, including Claude 2, Gemini Pro, and GPT4 0613. We show some selected models
from the leaderboard in the table. We note that many of those competing models contain
either proprietary alignment data (which is typically large, e.g., over 1M annotations in
[Touvron et al., 2023]) or use targets that are distilled from stronger models. In contrast,
our Self-Rewarding model starts from a small set of seed data from Open Assistant, and
then generates targ |
ets and rewards from the model itself for further iterations of training.
Preference optimization outperforms augmenting with positive examples only
We also tried the alternative self-training procedure of adding high-quality self-instruction
creation examples to supervised fine-tuning (without preference optimization). Unfortunately
we could not find a configuration where this approach helped. For example, adding 11,254
such examples that scored 5 out of 5, and optimizing the mixing weight in training, sti |
ll
yielded a head to head with the SFT Baseline of 29% wins vs 30% wins, i.e., no improvement.
6
Self-Rewarding M_3
vs.
SFT Baseline
Self-Rewarding M_2
vs.
SFT Baseline
Self-Rewarding M_1
vs.
SFT Baseline
62.5
49.2
30.5
27.7
36.3
38.7
9.8
14.5
30.9
Self-Rewarding Wins
Tie
SFT Baseline Wins
Self-Rewarding M_3
vs.
M_2
Self-Rewarding M_2
vs.
M_1
Self-Rewarding M_3
vs.
M_1
47.7
55.5
68.8
39.8
32.8
22.7
12.5
11.7
8.6
Left Wins (in Left vs. Right)
Tie
Right Wins
Figure 3: Instruction following ability improves wi |
th Self-Training: We evaluate our
models using head-to-head win rates on diverse prompts using GPT-4. The SFT Baseline is
on par with Self-Rewarding Iteration 1 (M1). However, Iteration 2 (M2) outperforms both
Iteration 1 (M1) and the SFT Baseline. Iteration 3 (M3) gives further gains over Iteration 2
(M2), outperforming M1, M2 and the SFT Baseline by a large margin.
Data distribution analysis
We perform a t-SNE [Van der Maaten and Hinton, 2008]
visualization of the IFT, EFT and AIFT(M1) data, shown in Sect |
ion A.1. We find good
overlap between the IFT and AIFT(M1) examples, which is desired, while the EFT examples
lie in a different part of the embedding space. We observe that generations from M1 have
an average length of 1092, for M2 they are 1552, and for M3 they are 2552, so the model is
learning to generate longer responses, which we note may be a factor in relative performance.
3.2.2
Reward Modeling Ability
Reward modeling evaluation results are provided in Table 2.
EFT augmentation improves over SFT bas |
eline
Firstly, we find that adding Eval-
uation Fine-Tuning (EFT) data into training, which gives examples to the model of how
to act as an LLM-as-a-Judge, naturally improves its performance compared to training
with Instruction Fine-Tuning (IFT) data alone. IFT data covers a wide range of general
instruction tasks, and so does endow the SFT Baseline with the ability to evaluate responses,
however EFT data gives more examples of this specific task. We find improvements across
all five metrics measured when |
using IFT+EFT vs. IFT alone, e.g. the pairwise accuracy
agreement with humans increases from 65.1% to 78.7%.
Reward Modeling ability improves with Self-Training
We find that performing
a round of self-reward training improves the ability of the model at providing self-rewards
for the next iteration, in addition to its improved instruction following ability. Model M2
(Iteration 2) is trained using the reward model from M1 (Iteration 1), but provides improved
performance on all five metrics compared to M1. Fo |
r example, pairwise accuracy improves
from 78.7% to 80.4%. Iteration 3 (M3) improves several of these metrics further compared
7
Table 1: AlpacaEval 2.0 results (win rate over GPT-4 Turbo evaluated by GPT-4).
Self-Rewarding iterations yield improving win rates. Iteration 3 (M3) outperforms many
existing models that use proprietary training data or targets distilled from stronger models.
Alignment Targets
Model
Win Rate
Distilled
Proprietary
Self-Rewarding 70B
Iteration 1 (M1)
9.94%
Iteration 2 (M2)
15.38%
I |
teration 3 (M3)
20.44%
Selected models from the leaderboard
GPT-4 0314
22.07%
✓
Mistral Medium
21.86%
✓
Claude 2
17.19%
✓
Gemini Pro
16.85%
✓
GPT-4 0613
15.76%
✓
GPT 3.5 Turbo 0613
14.13%
✓
LLaMA2 Chat 70B
13.87%
✓
Vicuna 33B v1.3
12.71%
✓
Humpback LLaMa2 70B
10.12%
Guanaco 65B
6.86%
Davinci001
2.76%
✓
Alpaca 7B
2.59%
✓
Table 2: Reward Modeling ability improves with Self-Training: We evaluate the LLM-
as-a-Judge via various metrics which measure alignment with held-out human preference data.
Self-Rewarding |
Iteration 2 (Model M2), which is trained using the self-reward model derived
from its previous iteration M1 outperforms Iteration 1 (M1), while M1 itself outperforms a
standard SFT baseline model trained on only Instruction Fine-Tuning (IFT) data. Iteration
3 (Model M3) gives further improvements over Iteration 2.
Self-Rewarding Models
Model
SFT Baseline
Iter 1 (M1)
Iter 2 (M2)
Iter 3 (M3)
Training data
IFT
IFT+EFT
IFT+EFT
IFT+EFT
+AIFT(M1)
+AIFT(M2)
Pairwise accuracy (↑)
65.1%
78.7%
80.4%
81.7%
5-best % (↑ |
)
39.6%
41.5%
44.3%
43.2%
Exact Match % (↑)
10.1%
13.1%
14.3%
14.3%
Spearman corr. (↑)
0.253
0.279
0.331
0.349
Kendall τ corr. (↑)
0.233
0.253
0.315
0.324
to M2, for example pairwise accuracy increases from 80.4% to 81.7%. This performance gain
is achieved despite there being no additional EFT data provided, and the examples created
during the Self-Instruction creation loop do not tend to look like LLM-as-a-Judge training
examples. We hypothesize that because the model is becoming better at general instruct |
ion
following, it nevertheless also improves at the LLM-as-a-Judge task.
Importance of the LLM-as-a-Judge Prompt
In these experiments we used the LLM-
as-Judge prompt format shown in Figure 2. In preliminary experiments we also tried various
other prompts to decide the most effective one to use. For example, we tried the prompt
proposed in Li et al. [2023a] which also proposes a 5-point scale, but describes the options as
multiple choice in a range of quality buckets, see Figure 5. In contrast, our prompt d |
escribes
the points as additive, covering various aspects of quality. We find a large difference between
8
these two prompts when using the SFT Baseline, e.g. 65.1% pairwise accuracy for ours, and
only 26.6% pairwise accuracy for theirs. See Section A.2 for further details.
4
Related Work
Automatically improving or self-correcting large language models is becoming a major focus
of research. A recent survey from Pan et al. [2023] attempts to summarize the topic. However,
this is a rapidly moving area, and th |
ere are already promising new works not covered there.
Reinforcement Learning from Human Feedback (RLHF)
Preference learning ap-
proaches such as in Ziegler et al. [2019], Stiennon et al. [2020], Ouyang et al. [2022], Bai et al.
[2022a] train a fixed reward model from human preference data, and then use the reward
model to train via reinforcement learning (RL), e.g. via Proximal Policy Optimization (PPO)
[Schulman et al., 2017]. Thus, the reward signal in a certain sense already comes from a model
even in t |
hese works, but distilled from human data. Nevertheless, this is commonly referred
to as RL from Human Feedback (RLHF). Methods such as Direct Preference Optimization
(DPO) [Rafailov et al., 2023] avoid training the reward model entirely, and instead directly
train the LLM using human preferences. Several other such competing methods exist as
well [Zhao et al., 2023, Zheng et al., 2023a, Yuan et al., 2023], including Pairwise Cringe
Optimization (PCO) [Xu et al., 2023], which was shown to outperform DPO and |
PPO on
AlpacaFarm [Dubois et al., 2023]. PCO uses an iterative training approach similar to the
one in our work, except with a fixed reward model, and also showed that Iterative DPO
improves over DPO using the same scheme.
Reinforcement Learning from AI Feedback (RLAIF)
Constitutional AI [Bai et al.,
2022b] uses an LLM to give feedback and refine responses, and uses this data to train a
(separate, fixed) reward model. This is then used to perform “RL from AI Feedback” (RLAIF).
Lee et al. [2023] compare RLA |
IF and RLHF procedures and find the methods they compare
perform roughly equally. They use an “off-the-shelf” LLM to perform LLM-as-a-Judge
prompting to build a training set to train a fixed reward model, which is then used for RL
training. They also experiment with using the fixed but separate LLM-as-a-Judge model
directly, which the authors report is computationally expensive due to using it within PPO
training (rather than the offline step in the iterative approach we use in our work, which
is relatively |
computationally cheap). Finally, Chen et al. [2024] recently showed they can
avoid reward models entirely in an Iterative DPO-like framework by using human labels as
the winning response in a pair, and the last iteration’s generations as the losing response in
the pair. The authors note this has the limitation that once the model generations reach
human performance, they are bottlenecked. Further, each input prompt is required to have
a human annotated response, in contrast to our work.
Improving LLMs via |
data augmentation (and curation)
Several methods have
improved LLMs by (self-)creating training data to augment fine-tuning. Self-Instruct [Wang
et al., 2022] is a method for self-instruction creation of prompts and responses, which can
be used to improve a base LLM. We make use of a similar technique in our work, and then
use our self-reward model to score them. Several approaches have also created training data
by distilling from powerful LLMs, and shown a weaker LLM can then perform well. For
example, Al |
paca [Taori et al., 2023] fine-tuned a Llama 7B model with text-davinci-003
instructions created in the style of self-instruct. Alpagasus [Chen et al., 2023] employed a
strong LLM-as-a-Judge (ChatGPT) to curate the Alpaca dataset and filter to a smaller set,
obtaining improved results. Instruction Backtranslation [Li et al., 2023a] similarly augments
and curates training data, but augmenting via backtranslating from web documents to predict
prompts. The curation is done by the LLM(-as-a-Judge) itself, so ca |
n be seen as an instance
of a self-rewarding model, but in a specialized setting. Reinforced Self-Training (ReST)
[Gulcehre et al., 2023] uses a fixed, external reward to curate new high-quality examples to
iteratively add to the training set, improving performance. We note that in our experiments,
we found that adding only positive examples in a related manner did not help, whereas
adding preference pairs did help.
9
LLM-as-a-Judge
Using LLM-as-a-Judge prompting to evaluate language models has
become a sta |
ndard approach [Dubois et al., 2023, Li et al., 2023b, Fernandes et al., 2023,
Bai et al., 2023, Saha et al., 2023], and is being used to train reward models or curate data
as well, as described above [Lee et al., 2023, Chen et al., 2023, Li et al., 2023a]. While some
works such as Kim et al. [2023] create training to data to train an LLM to perform well as a
judge, to our knowledge it is not common to combine this training with general instruction
following skills as in our work.
5
Conclusion
We have intro |
duced Self-Rewarding Language Models, models capable of self-alignment via
judging and training on their own generations. The method is trained in an iterative manner,
where in each iteration the model creates its own preference-based instruction training data.
This is done by assigning rewards to its own generations via LLM-as-a-Judge prompting,
and using Iterative DPO to train on the preferences. We showed that this training both
improves the instruction following capability of the model, as well as its r |
eward-modeling
ability across the iterations. While this is only a preliminary study, we believe this is an
exciting avenue of research because this means the model is better able to assign rewards
in future iterations for improving instruction following – a kind of virtuous circle. While
this improvement likely saturates in realistic scenarios, it still allows for the possibility of
continual improvement beyond the human preferences that are typically used to build reward
models and instruction following m |
odels today.
6
Limitations
While we have obtained promising experimental results, we currently consider them pre-
liminary because there are many avenues yet to explore, among them the topics of further
evaluation, including safety evaluation, and understanding the limits of iterative training.
We showed that the iterations of training improve both instruction following and reward
modeling ability, but only ran three iterations in a single setting. A clear line of further
research is to understand the “scal |
ing laws” of this effect both for more iterations, and with
different language models with more or less capabilities in different settings.
While we have evaluated our models using GPT-4 using head-to-head and AlpacaEval 2
leaderboard style evaluation, there are many other automatic evaluation benchmarks that
one can measure. Further, we observed an increase in length in model generations, and there
is a known correlation between length and estimated quality, which is a topic that should be
understood more |
deeply in general, and in our results in particular as well. It would also be
good to understand if so-called “’reward-hacking” can happen within our framework, and in
what circumstances. As we are using both a language model as the training reward, and
a language model for final evaluation, even if they are different models, this may require
a deeper analysis than we have provided. We conducted a preliminary human (author)
evaluation, which validated the automatic results that we see, but more detailed hum |
an
evaluation would be beneficial.
Another clear further avenue of study is to conduct safety evaluations – and to explore safety
training within our framework. Reward models have been built exclusively for safety in
existing systems [Touvron et al., 2023], and a promising avenue here would be to use the
LLM-as-a-Judge procedure to evaluate for safety specifically in our self-rewarding training
process. Given that we have shown that reward modeling ability improves over training
iterations, this could mean |
that the safety of the model could potentially improve over time
as well, with later iterations being able to capture more challenging safety situations that
earlier iterations cannot.
10
References
Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni
Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al.
GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2023.
Leonard Adolphs, Tianyu Gao, Jing Xu, Kurt Shuster, Sainbayar Sukhbaatar, and Jason
W |
eston. The CRINGE loss: Learning what language not to model. In Anna Rogers, Jordan
Boyd-Graber, and Naoaki Okazaki, editors, Proceedings of the 61st Annual Meeting of
the Association for Computational Linguistics (Volume 1: Long Papers), pages 8854–8874,
Toronto, Canada, July 2023. Association for Computational Linguistics. doi: 10.18653/
v1/2023.acl-long.493. URL https://aclanthology.org/2023.acl-long.493.
Anthropic. Claude 2. https://www.anthropic.com/index/claude-2, 2023.
Yuntao Bai, Andy Jones, Kamal N |
dousse, Amanda Askell, Anna Chen, Nova DasSarma,
Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful
and harmless assistant with reinforcement learning from human feedback. arXiv preprint
arXiv:2204.05862, 2022a.
Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy
Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitu-
tional AI: Harmlessness from AI feedback. arXiv preprint arXiv:2212.08073, 2022b.
Yushi Bai, Jiahao Yi |
ng, Yixin Cao, Xin Lv, Yuze He, Xiaozhi Wang, Jifan Yu, Kaisheng
Zeng, Yijia Xiao, Haozhe Lyu, Jiayin Zhang, Juanzi Li, and Lei Hou. Benchmarking
foundation models with language-model-as-an-examiner. In Thirty-seventh Conference
on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. URL
https://openreview.net/forum?id=IiRHQ7gvnq.
Lichang Chen, Shiyang Li, Jun Yan, Hai Wang, Kalpa Gunaratna, Vikas Yadav, Zheng Tang,
Vijay Srinivasan, Tianyi Zhou, Heng Huang, et al. AlpaGasus: Training |
a better alpaca
with fewer data. arXiv preprint arXiv:2307.08701, 2023.
Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning
converts weak language models to strong language models. arXiv preprint arXiv:2401.01335,
2024.
Ronan Collobert and Jason Weston. A unified architecture for natural language processing:
Deep neural networks with multitask learning. In Proceedings of the 25th International
Conference on Machine Learning, pages 160–167, 2008.
Yann Dubois, Xuechen Li, |
Rohan Taori, Tianyi Zhang, Ishaan Gulrajani, Jimmy Ba, Carlos
Guestrin, Percy Liang, and Tatsunori B Hashimoto. Alpacafarm: A simulation framework
for methods that learn from human feedback. arXiv preprint arXiv:2305.14387, 2023.
Patrick Fernandes, Daniel Deutsch, Mara Finkelstein, Parker Riley, André FT Martins,
Graham Neubig, Ankush Garg, Jonathan H Clark, Markus Freitag, and Orhan Firat.
The devil is in the errors: Leveraging large language models for fine-grained machine
translation evaluation. arXiv p |
reprint arXiv:2308.07286, 2023.
Caglar Gulcehre, Tom Le Paine, Srivatsan Srinivasan, Ksenia Konyushkova, Lotte Weerts,
Abhishek Sharma, Aditya Siddhant, Alex Ahern, Miaosen Wang, Chenjie Gu, et al.
Reinforced self-training (rest) for language modeling. arXiv preprint arXiv:2308.08998,
2023.
Or Honovich, Thomas Scialom, Omer Levy, and Timo Schick. Unnatural instructions: Tuning
language models with (almost) no human labor. In Anna Rogers, Jordan Boyd-Graber,
and Naoaki Okazaki, editors, Proceedings of the 61 |
st Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long Papers), pages 14409–14428, Toronto, Canada,
July 2023. Association for Computational Linguistics. doi: 10.18653/v1/2023.acl-long.806.
URL https://aclanthology.org/2023.acl-long.806.
11
Seungone Kim, Jamin Shin, Yejin Cho, Joel Jang, Shayne Longpre, Hwaran Lee, Sangdoo Yun,
Seongjin Shin, Sungdong Kim, James Thorne, et al. Prometheus: Inducing fine-grained
evaluation capability in language models. arXiv preprint arXiv:2310.08 |
491, 2023.
Andreas Köpf, Yannic Kilcher, Dimitri von Rütte, Sotiris Anagnostidis, Zhi-Rui Tam,
Keith Stevens, Abdullah Barhoum, Nguyen Minh Duc, Oliver Stanley, Richárd Nagyfi,
et al. OpenAssistant conversations–democratizing large language model alignment. arXiv
preprint arXiv:2304.07327, 2023.
Harrison Lee, Samrat Phatale, Hassan Mansoor, Kellie Lu, Thomas Mesnard, Colton Bishop,
Victor Carbune, and Abhinav Rastogi. RLAIF: Scaling reinforcement learning from human
feedback with ai feedback. arXiv preprint |
arXiv:2309.00267, 2023.
Xian Li, Ping Yu, Chunting Zhou, Timo Schick, Luke Zettlemoyer, Omer Levy, Jason
Weston, and Mike Lewis. Self-alignment with instruction backtranslation. arXiv preprint
arXiv:2308.06259, 2023a.
Xuechen Li, Tianyi Zhang, Yann Dubois, Rohan Taori, Ishaan Gulrajani, Carlos Guestrin,
Percy Liang, and Tatsunori B. Hashimoto.
Alpacaeval: An automatic evaluator of
instruction-following models. https://github.com/tatsu-lab/alpaca_eval, 2023b.
Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida |
, Carroll Wainwright, Pamela Mishkin,
Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language
models to follow instructions with human feedback. Advances in Neural Information
Processing Systems, 35:27730–27744, 2022.
Liangming Pan, Michael Saxon, Wenda Xu, Deepak Nathani, Xinyi Wang, and William Yang
Wang. Automatically correcting large language models: Surveying the landscape of diverse
self-correction strategies. arXiv preprint arXiv:2308.03188, 2023.
Alec Radford, Jeffrey Wu, R |
ewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al.
Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019.
Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and
Chelsea Finn. Direct preference optimization: Your language model is secretly a reward
model. In Thirty-seventh Conference on Neural Information Processing Systems, 2023.
URL https://openreview.net/forum?id=HPuSIXJaa9.
Swarnadeep Saha, Omer Levy, Asli Celikyilmaz, Mohit Bansal, Jas |
on Weston, and Xian
Li. Branch-solve-merge improves large language model evaluation and generation. arXiv
preprint arXiv:2310.15123, 2023.
John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal
policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.
Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec
Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human
feedback. Advances in Neural Information Process |
ing Systems, 33:3008–3021, 2020.
Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin,
Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following
llama model. https://github.com/tatsu-lab/stanford_alpaca, 2023.
Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei,
Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2:
Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09 |
288, 2023.
Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of
machine learning research, 9(11), 2008.
Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi,
and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated
instructions. arXiv preprint arXiv:2212.10560, 2022.
12
Jing Xu, Andrew Lee, Sainbayar Sukhbaatar, and Jason Weston. Some things are more
cringe than others: Preference optimization with the pairwise cring |
e loss. arXiv preprint
arXiv:2312.16682, 2023.
Hongyi Yuan, Zheng Yuan, Chuanqi Tan, Wei Wang, Songfang Huang, and Fei Huang.
RRHF: Rank responses to align language models with human feedback. In Thirty-seventh
Conference on Neural Information Processing Systems, 2023. URL https://openreview.
net/forum?id=EdIGMCHk4l.
Yao Zhao, Rishabh Joshi, Tianqi Liu, Misha Khalman, Mohammad Saleh, and Peter J
Liu. SLiC-HF: Sequence likelihood calibration with human feedback. arXiv preprint
arXiv:2305.10425, 2023.
Chujie |
Zheng, Pei Ke, Zheng Zhang, and Minlie Huang. Click: Controllable text generation
with sequence likelihood contrastive learning. In Anna Rogers, Jordan Boyd-Graber, and
Naoaki Okazaki, editors, Findings of the Association for Computational Linguistics: ACL
2023, pages 1022–1040, Toronto, Canada, July 2023a. Association for Computational
Linguistics. doi: 10.18653/v1/2023.findings-acl.65. URL https://aclanthology.org/
2023.findings-acl.65.
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu |
, Yonghao
Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, Hao Zhang, Joseph E. Gonzalez, and
Ion Stoica. Judging LLM-as-a-judge with MT-bench and chatbot arena. In Thirty-seventh
Conference on Neural Information Processing Systems Datasets and Benchmarks Track,
2023b. URL https://openreview.net/forum?id=uccHPGDlao.
Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei,
Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences.
arXiv preprint |
arXiv:1909.08593, 2019.
13
A
Appendix
A.1
Distributions of IFT, EFT and AIFT data
100
50
0
50
Dimension 1
75
50
25
0
25
50
75
Dimension 2
IFT data
EFT data
AIFT data
(a) Instruction distribution of IFT, EFT and
AIFT data.
50
0
50
Dimension 1
80
60
40
20
0
20
40
60
Dimension 2
IFT data
EFT data
AIFT data
(b) Response distribution of IFT, EFT, and AIFT
data.
Figure 4: Distributions of both instructions and responses for IFT, EFT and AIFT data.
We have plotted the distribution of instructions for IFT, EFT and |
AIFT data, and the
distribution of responses for IFT, EFT and AIFT data in Figure 4. It is clear that the IFT
data and EFT data come from very different distributions while the IFT and AIFT data
come from similar distributions.
A.2
Other EFT prompts we have tried
At first, we took the EFT prompt from Li et al. [2023a] as shown in Figure 5. However, we
found that this prompt was not as effective as our additive score-counting prompt because
the model needed to treat the task as a multiple-choice problem, and |
it was difficult for the
model to break down this multiple-choice problem into sub-problems involving evaluating
various aspects of the response. When using the model trained on 3,200 IFT data only, its
performance on the EFT test set using our additive score-counting prompt and prompt from
Li et al. [2023a] is shown in Table 3.
EFT Prompt
Multiple Choice prompt
Ours
Pairwise accuracy (↑)
26.6%
65.1%
5-best % (↑)
23.5%
39.6%
Exact Match % (↑)
1.1%
10.1%
Spearman corr. (↑)
-0.18
0.25
Kendall τ corr. (↑)
-0. |
16
0.23
Table 3: We tried various LLM-as-Judge prompts using the model trained with 3,200 IFT
data only and found that our additive score-counting prompt worked best which demonstrates
significant improvements in EFT performance comparing to the prompt used by Li et al.
[2023a].
14
Below is a question from an user and a candidate response. Please grade the response on a
5-point scale using the following criteria:
1:
It means the answer is incomplete, vague, off-topic, controversial, or not ex-
actly what th |
e user asked for. For example, some content seems missing, numbered list
does not start from the beginning, the opening sentence repeats user’s question. Or the
response is from another person’s perspective with their personal experience (e.g. taken
from blog posts), or looks like an answer from a forum. Or it contains promotional text,
navigation text, or other irrelevant information.
2: It means the answer addresses most of the asks from the user. It does not directly
address the user’s question. For exam |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.