介绍

本模型主要用途为基于科技类文章生成对应标题。

本次将开源从 100-2200 steps 的中间所有 checkpoint 以供大家参考。

使用

from transformers import AutoModelForCausalLM, AutoTokenizer,BitsAndBytesConfig

peft_model_id = "checkpoint-2000"
model = AutoModelForCausalLM.from_pretrained(peft_model_id,device_map="cuda")

tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")

input_text = """
Generate a title for the article:

{content}

---
Title:
""" # 固定格式
encoding = tokenizer(input_text, return_tensors="pt").to("cuda")

outputs = model.generate(**encoding,max_length=8192,temperature=0.2,do_sample=True)
generated_ids = outputs[:, encoding.input_ids.shape[1]:]
generated_texts = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
print(generated_texts[0])

训练数据

linux-cn 文章 https://huggingface.co/datasets/linux-cn/archive

微调

基于 LLaMA-Factory 在单张A100(80G)上进行微调,微调参数如下:

CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
    --stage sft \
    --do_train True \
    --model_name_or_path google/gemma-2b \
    --finetuning_type lora \
    --template default \
    --dataset title \
    --use_unsloth \
    --cutoff_len 8192 \
    --learning_rate 5e-05 \
    --num_train_epochs 10.0 \
    --max_samples 10000 \
    --per_device_train_batch_size 4 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --max_grad_norm 1.0 \
    --logging_steps 10 \
    --save_steps 100 \
    --eval_steps 100 \
    --evaluation_strategy steps \
    --warmup_steps 0 \
    --output_dir saves/Gemma-2B/lora/train_2024-03-01-04-36-32 \
    --bf16 True \
    --lora_rank 8 \
    --lora_dropout 0.1 \
    --lora_target q_proj,v_proj \
    --val_size 0.1 \
    --load_best_model_at_end True \
    --plot_loss True \
    --report_to "tensorboard"

如果你需要查看详细的流程,可以查看如下文章

Google Gemma 2B 微调实战(IT科技新闻标题生成)

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Examples
Unable to determine this model's library. Check the docs .

Dataset used to train GanymedeNil/gemma-2b-technology-news-title-generation-lora