|
--- |
|
license: apache-2.0 |
|
datasets: |
|
- tatsu-lab/alpaca |
|
- sahil2801/CodeAlpaca-20k |
|
language: |
|
- zh |
|
- en |
|
library_name: transformers |
|
tags: |
|
- baichuan |
|
- lora |
|
--- |
|
|
|
A bilingual instruction-tuned LoRA model of https://huggingface.co/baichuan-inc/baichuan-7B |
|
|
|
- Instruction-following datasets used: alpaca, alpaca-zh, codealpaca |
|
- Training framework: https://github.com/hiyouga/LLaMA-Efficient-Tuning |
|
|
|
Usage: |
|
|
|
```python |
|
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer |
|
|
|
tokenizer = AutoTokenizer.from_pretrained("hiyouga/baichuan-7b-sft", trust_remote_code=True) |
|
model = AutoModelForCausalLM.from_pretrained("hiyouga/baichuan-7b-sft", trust_remote_code=True).cuda() |
|
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True) |
|
|
|
query = "晚上睡不着怎么办" |
|
template = "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.\nHuman: {}\nAssistant: " |
|
|
|
inputs = tokenizer([template.format(query)], return_tensors="pt") |
|
inputs = inputs.to("cuda") |
|
generate_ids = model.generate(**inputs, max_new_tokens=256, streamer=streamer) |
|
``` |
|
|
|
You could also alternatively launch a CLI demo by using the script in https://github.com/hiyouga/LLaMA-Efficient-Tuning |
|
|
|
```bash |
|
python src/cli_demo.py --model_name_or_path hiyouga/baichuan-7b-sft |
|
``` |
|
|
|
--- |
|
|
|
You could reproduce our results with the following scripts: |
|
|
|
```bash |
|
CUDA_VISIBLE_DEVICES=0 python src/train_sft.py \ |
|
--model_name_or_path baichuan-inc/baichuan-7B \ |
|
--do_train \ |
|
--dataset alpaca_gpt4_en,alpaca_gpt4_zh,codealpaca \ |
|
--finetuning_type lora \ |
|
--lora_rank 16 \ |
|
--lora_target W_pack,o_proj,gate_proj,down_proj,up_proj \ |
|
--output_dir baichuan_lora \ |
|
--overwrite_cache \ |
|
--per_device_train_batch_size 8 \ |
|
--per_device_eval_batch_size 8 \ |
|
--gradient_accumulation_steps 8 \ |
|
--preprocessing_num_workers 16 \ |
|
--lr_scheduler_type cosine \ |
|
--logging_steps 10 \ |
|
--save_steps 100 \ |
|
--eval_steps 100 \ |
|
--learning_rate 5e-5 \ |
|
--max_grad_norm 0.5 \ |
|
--num_train_epochs 2.0 \ |
|
--dev_ratio 0.01 \ |
|
--evaluation_strategy steps \ |
|
--load_best_model_at_end \ |
|
--plot_loss \ |
|
--fp16 |
|
``` |
|
|
|
Loss curve on training set: |
|
 |
|
|
|
Loss curve on evaluation set: |
|
 |
