ShinojiResearch/Senku-70B-Full

UPDATE: 85.09 EQ-Bench with ChatML teamplate

• EQ-Bench: (Mistral) 84.89 -> 85.09 (ChatML)
• GSM8k: (Mistral) 77.18 -> 71.04 (ChatML)
• Hellaswag: (Mistral) 87.67 -> ??

Finetune of miqu-70b-sf dequant of miqudev's leak of Mistral-70B (allegedly an early mistral medium). My diffs are available under CC-0 (That is the Senku-70B repo, full includes the merge), this is a merge with the leaked model, you can use the other repository to save bandwidth.

Update: Upon further testing a score of 85.09 was achieved using ChatML instead of Mistral's prompt.

Prompt Template

I recommend using the ChatML format instead, I will run more benchmarks. This also fixes the bug with Miqu dequant failing to provide a stop.

<|im_start|>system 
Provide some context and/or instructions to the model.
<|im_end|> 
<|im_start|>user 
The user’s message goes here
<|im_end|> 
<|im_start|>assistant <|im_end|>

Kudos

Credit to https://twitter.com/hu_yifei for providing GSM & Hellaswag. It is the first open weight model to dethrone GPT-4 on EQ bench.

Base Model Details

This model is a fine-tuned version of 152334H/miqu-1-70b-sf on the Slimorca dataset. It achieves the following results on the evaluation set:

• Loss: 0.3110

Training procedure

See axolotl config

axolotl version: 0.4.0

base_model: 152334H/miqu-1-70b-sf
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer
is_mistral_derived_model: true

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: Open-Orca/SlimOrca
    type: sharegpt
    conversation: chatml
dataset_prepared_path: last_run_prepared
val_set_size: 0.1
output_dir: ./qlora-out

adapter: qlora
lora_model_dir:

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 1
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3

warmup_steps: 10
evals_per_epoch: 4
eval_table_size:
eval_table_max_new_tokens: 128
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
0.9043	0.0	1	0.6387
0.5612	0.25	881	0.3279
0.6044	0.5	1762	0.3177
0.6592	0.75	2643	0.3110

Framework versions

• PEFT 0.8.2
• Transformers 4.38.0.dev0
• Pytorch 2.1.2+cu118
• Datasets 2.16.1
• Tokenizers 0.15.0

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	75.44
AI2 Reasoning Challenge (25-Shot)	71.50
HellaSwag (10-Shot)	87.88
MMLU (5-Shot)	75.20
TruthfulQA (0-shot)	61.96
Winogrande (5-shot)	84.77
GSM8k (5-shot)	71.34