mlabonne/NeuralHermes-2.5-Mistral-7B-laser

NeuralHermes 2.5 - Mistral 7B - LASER

This is an experimental LASER version of NeuralHermes using laserRMT, based on this paper.

Model	AGIEval	GPT4All	TruthfulQA	Bigbench	Average
NeuralHermes-2.5-Mistral-7B-laser	43.54	73.44	55.26	42.24	53.62
NeuralHermes-2.5-Mistral-7B	43.67	73.24	55.37	41.76	53.51

Fernando Fernandes Neto and Eric Hartford. "Optimizing Large Language Models Using Layer-Selective Rank Reduction and Random Matrix Theory." 2024.

NeuralHermes is an teknium/OpenHermes-2.5-Mistral-7B model that has been further fine-tuned with Direct Preference Optimization (DPO) using the mlabonne/chatml_dpo_pairs dataset. It surpasses the original model on several benchmarks (see results).

It is directly inspired by the RLHF process described by Intel/neural-chat-7b-v3-1's authors to improve performance. I used the same dataset and reformatted it to apply the ChatML template.

The code to train this model is available on Google Colab and GitHub. It required an A100 GPU for about an hour.

Results

AGIEval

Task	Version	Metric	Value		Stderr
agieval_aqua_rat	0	acc	21.26	±	2.57
		acc_norm	22.83	±	2.64
agieval_logiqa_en	0	acc	39.32	±	1.92
		acc_norm	40.71	±	1.93
agieval_lsat_ar	0	acc	25.65	±	2.89
		acc_norm	25.65	±	2.89
agieval_lsat_lr	0	acc	48.82	±	2.22
		acc_norm	50.00	±	2.22
agieval_lsat_rc	0	acc	58.36	±	3.01
		acc_norm	57.25	±	3.02
agieval_sat_en	0	acc	74.27	±	3.05
		acc_norm	73.30	±	3.09
agieval_sat_en_without_passage	0	acc	43.69	±	3.46
		acc_norm	42.23	±	3.45
agieval_sat_math	0	acc	37.27	±	3.27
		acc_norm	36.36	±	3.25

Average: 43.54%

GPT4All

Task	Version	Metric	Value		Stderr
arc_challenge	0	acc	57.76	±	1.44
		acc_norm	60.32	±	1.43
arc_easy	0	acc	83.84	±	0.76
		acc_norm	81.10	±	0.80
boolq	1	acc	86.70	±	0.59
hellaswag	0	acc	63.15	±	0.48
		acc_norm	82.55	±	0.38
openbookqa	0	acc	34.40	±	2.13
		acc_norm	45.20	±	2.23
piqa	0	acc	81.94	±	0.90
		acc_norm	82.97	±	0.88
winogrande	0	acc	75.22	±	1.21

Average: 73.44%

TruthfulQA

Task	Version	Metric	Value		Stderr
truthfulqa_mc	1	mc1	37.70	±	1.70
		mc2	55.26	±	1.52

Average: 55.26%

Bigbench

Task	Version	Metric	Value		Stderr
bigbench_causal_judgement	0	multiple_choice_grade	53.16	±	3.63
bigbench_date_understanding	0	multiple_choice_grade	65.31	±	2.48
bigbench_disambiguation_qa	0	multiple_choice_grade	34.11	±	2.96
bigbench_geometric_shapes	0	multiple_choice_grade	27.02	±	2.35
		exact_str_match	0.28	±	0.28
bigbench_logical_deduction_five_objects	0	multiple_choice_grade	27.80	±	2.01
bigbench_logical_deduction_seven_objects	0	multiple_choice_grade	19.86	±	1.51
bigbench_logical_deduction_three_objects	0	multiple_choice_grade	48.33	±	2.89
bigbench_movie_recommendation	0	multiple_choice_grade	41.40	±	2.20
bigbench_navigate	0	multiple_choice_grade	50.00	±	1.58
bigbench_reasoning_about_colored_objects	0	multiple_choice_grade	65.00	±	1.07
bigbench_ruin_names	0	multiple_choice_grade	46.21	±	2.36
bigbench_salient_translation_error_detection	0	multiple_choice_grade	27.25	±	1.41
bigbench_snarks	0	multiple_choice_grade	70.72	±	3.39
bigbench_sports_understanding	0	multiple_choice_grade	65.72	±	1.51
bigbench_temporal_sequences	0	multiple_choice_grade	30.40	±	1.46
bigbench_tracking_shuffled_objects_five_objects	0	multiple_choice_grade	22.56	±	1.18
bigbench_tracking_shuffled_objects_seven_objects	0	multiple_choice_grade	17.09	±	0.90
bigbench_tracking_shuffled_objects_three_objects	0	multiple_choice_grade	48.33	±	2.89

Average: 42.24%

Average score: 53.62%

Usage

You can run this model using LM Studio or any other frontend.

You can also run this model using the following code:

import transformers
from transformers import AutoTokenizer

# Format prompt
message = [
    {"role": "system", "content": "You are a helpful assistant chatbot."},
    {"role": "user", "content": "What is a Large Language Model?"}
]
tokenizer = AutoTokenizer.from_pretrained(new_model)
prompt = tokenizer.apply_chat_template(message, add_generation_prompt=True, tokenize=False)

# Create pipeline
pipeline = transformers.pipeline(
    "text-generation",
    model="mlabonne/NeuralHermes-2.5-Mistral-7B-laser",
    tokenizer=tokenizer
)

# Generate text
sequences = pipeline(
    prompt,
    do_sample=True,
    temperature=0.7,
    top_p=0.9,
    num_return_sequences=1,
    max_length=200,
)
print(sequences[0]['generated_text'])

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	67.29
AI2 Reasoning Challenge (25-Shot)	66.38
HellaSwag (10-Shot)	85.09
MMLU (5-Shot)	63.43
TruthfulQA (0-shot)	54.95
Winogrande (5-shot)	78.14
GSM8k (5-shot)	55.72