modelscope/llama3-8b-agent-instruct-v2

Fine-tuning the llama3-8b-instruct model using the msagent-pro dataset and the loss_scale technique with swift, the script is as follows:

NPROC_PER_NODE=8 \
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
MASTER_PORT=29500 \
swift sft \
    --model_type llama3-8b-instruct \
    --learning_rate 2e-5 \
    --sft_type lora \
    --dataset msagent-pro \
    --gradient_checkpointing true \
    --gradient_accumulation_steps 8 \
    --deepspeed default-zero3 \
    --lora_target_modules ALL \
    --use_loss_scale true \
    --save_strategy epoch \
    --batch_size 1 \
    --num_train_epochs 2 \
    --max_length 4096 \
    --preprocess_num_proc 4 \
    --use_loss_scale true \
    --loss_scale_config_path agent-flan \
    --ddp_backend nccl \

Comparison with the Original Model on the ToolBench Evaluation Set

Model	ToolBench (in-domain)					ToolBench (out-of-domain)
	Plan.EM	Act.EM	HalluRate (lower is better)	Avg.F1	R-L	Plan.EM	Act.EM	HalluRate (lower is better)	Avg.F1
llama3-8b-instruct	74.22	36.17	15.68	20.0	12.14	69.47	34.21	14.72	20.25
llama3-8b-agent-instruct-v2	85.15	58.1	1.57	52.10	26.02	85.79	59.43	2.56	52.19

For detailed explanations of the evaluation metrics, please refer to document

Deploy this model:

USE_HF=True swift deploy \
  --model_id_or_path modelscope/llama3-8b-agent-instruct-v2 \
  --model_type llama3-8b-instruct \
  --infer_backend vllm \
  --tools_prompt toolbench