Issue with Fine-tuning Mistral 7B Model - Results Discrepancy

[lyf-00]

Hello,
I attempt to replicate the experiment using metamathQA dataset to finetune mistral-7b, but the results I obtained do not match the ones shared in the repository.

Reproduction steps

I used the following parameters in run_mistral.sh.

export MODEL_PATH='mistralai/Mistral-7B-v0.1'
export SAVE_PATH='0224_mistral-7b-metamath395'
export MASTER_ADDR="localhost"
export MASTER_PORT="1231"
export GLOO_SOCKET_IFNAME="lo"
export NCCL_SOCKET_IFNAME="lo"
export WANDB_DISABLED=true
export HF_TOKEN="token of your huggingface"
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 -m torch.distributed.launch --master_addr ${MASTER_ADDR} --master_port ${MASTER_PORT} --nproc_per_node=8 --use_env train_math.py \
    --model_name_or_path $MODEL_PATH \
    --data_path MetaMathQA-395K.json \
    --data_length 10000000 \
    --bf16 True \
    --output_dir $SAVE_PATH \
    --num_train_epochs 3 \
    --per_device_train_batch_size 2 \
    --per_device_eval_batch_size 2 \
    --gradient_accumulation_steps 8 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 100000 \
    --save_total_limit 0 \
    --learning_rate 5e-6 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --fsdp "full_shard auto_wrap" \
    --fsdp_transformer_layer_cls_to_wrap 'MistralDecoderLayer' \
    --tf32 True
python eval_gsm8k.py --model $SAVE_PATH --data_file ./data/test/GSM8K_test.jsonl
python eval_math.py --model $SAVE_PATH --data_file ./data/test/MATH_test.jsonl

and I get
gsm8k acc==== 0.6618650492797574
math acc==== 0.2274

which is different from the reported 77.7 and 28.2

Environment details

Here are the detailed of my Python environment:

transformers==4.34.0
wandb==0.15.3
torch==2.0.1
sentencepiece==0.1.99
tokenizers==0.14
accelerate==0.21.0
bitsandbytes==0.40.0

I would appreciate any guidance or suggestions you could provide to help resolve this discrepancy. Thank you in advance for your time and assistance.

Best regards,
lyf-00

[nuochenpku]

Same problem for me! Do you solve it? I also can not reproduce the results whatever the Mistral or LLaMA

[ytyz1307zzh]

I encountered the same issue when trying to train Mistral-7B on MetaMathQA. My environment is:

transformers==4.34.0
torch==2.0.1
sentencepiece==0.1.99
tokenizers==0.14.1
accelerate==0.21.0

I only got a 69% accuracy on GSM8K and 24% on MATH after 3 epochs with LR 5e-6 and global batch 128. Although due to the limitation of my computational resources, I added gradient checkpointing and flash attention to the original code, and also changed the per_device_batch_size to 1 (so gradient accumulates for 16 steps on 8 GPUs), but I don't think these modifications will bring significant differences to the performance.

[Nipers]

My result with llama-factory and hyperparameters reported in the paper is 72.2%on GSM8K, I do not under stand why there are so many failures when trying to reproduce the result.

[AaronZLT]

My result with llama-factory and hyperparameters reported in the paper is 72.2%on GSM8K, I do not under stand why there are so many failures when trying to reproduce the result.

I just can't get such a higher score, and I don't think this is an unique case