使用[Firefly](https://github.com/yangjianxin1/Firefly)项目微调baichuan-13b-base。训练数据约为一百万多轮对话数据，包括项目分享的moss数据+2万条school math数据。 更多详情见项目：[Firefly](https://github.com/yangjianxin1/Firefly) 技术细节分享：[Firefly增强Baichuan-13B的多轮对话能力](https://mp.weixin.qq.com/s/djO8Tg3emmy6wzw_rTUlcw) 训练loss： [//]: # () ![firefly_logo](firefly-baichuan-13b-loss.jpg) C-Eval榜单： | Model | C-Eval | STEM | Social Science | Humanities | Other | |----------------------------------|--------|-------|----------------|------------|-------| | Baichuan-13B-Chat(官方) | 52.05 | 42.23 | 65.27 | 58.61 | 51.32 | | **firefly-baichuan-13b** | 51.36 | 44.24 | 61.65 | 54.63 | 51.68 | | chatglm2-6b(官方) | 50.45 | 41.91 | 60.73 | 59.24 | 47.82 | | **firefly-chatglm2-6b** | 49.13 | 43.6 | 58.83 | 54.48 | 45.03 | | openbuddy-llama2-13b-v11.1-bf16 | 43.36 | 39.79 | 50.28 | 44.78 | 42.13 | | chinese-alpaca-2-13b(哈工大) | 41.86 | 36.52 | 49.7 | 47.97 | 38.33 | | openbuddy-llama2-13b-v8.1-fp16 | 41.62 | 38.82 | 44.66 | 40.28 | 45.32 | | chinese-alpaca-2-7b(哈工大) | 41.48 | 35.01 | 50.08 | 43.02 | 43.87 | | belle-llama2-13B-chat-0.4M | 41.11 | 40.04 | 44.71 | 42.09 | 38.82 | | ziya-llama-13b | 39.1 | - | - | - | - | | llama-2-13b-chat(官方) | 36.38 | 33.68 | 46.38 | 34.47 | 34.1 | | lama-2-7b-chat(官方) | 35.86 | 32.85 | 40.04 | 37.37 | 36.01 | | flagalpha/Llama2-Chinese-7b-Chat | 34.54 | 35.21 | 37.9 | 33.11 | 31.7 | | yayi-13b-llama2 | 34.15 | 36.48 | 30.64 | 32.67 | 34.6 | | yayi-7b-llama2 | 30.18 | 25.88 | 38.23 | 34.56 | 26.31 | | linly-llama2-7b | 28.35 | 26.06 | 33.47 | 29.71 | 26.53 | | linly-llama2-13b | 27.86 | 27.67 | 26.95 | 27.93 | 28.95 | 单轮对话: ```python from transformers import AutoModelForCausalLM, AutoTokenizer import torch """ 单轮对话，不具有对话历史的记忆功能 """ def main(): model_name = 'YeungNLP/firefly-baichuan-13b' max_new_tokens = 500 top_p = 0.9 temperature = 0.35 repetition_penalty = 1.0 device = 'cuda' model = AutoModelForCausalLM.from_pretrained( model_name, trust_remote_code=True, low_cpu_mem_usage=True, torch_dtype=torch.float16, device_map='auto' ).to(device).eval() tokenizer = AutoTokenizer.from_pretrained( model_name, trust_remote_code=True, # llama不支持fast use_fast=False if model.config.model_type == 'llama' else True ) # QWenTokenizer比较特殊，pad_token_id、bos_token_id、eos_token_id均为None。eod_id对应的token为<|endoftext|> if tokenizer.__class__.__name__ == 'QWenTokenizer': tokenizer.pad_token_id = tokenizer.eod_id tokenizer.bos_token_id = tokenizer.eod_id tokenizer.eos_token_id = tokenizer.eod_id text = input('User：') while True: text = text.strip() # chatglm使用官方的数据组织格式 if model.config.model_type == 'chatglm': text = '[Round 1]\n\n问：{}\n\n答：'.format(text) input_ids = tokenizer(text, return_tensors="pt", add_special_tokens=False).input_ids.to(device) # 为了兼容qwen-7b，因为其对eos_token进行tokenize，无法得到对应的eos_token_id else: input_ids = tokenizer(text, return_tensors="pt", add_special_tokens=False).input_ids.to(device) bos_token_id = torch.tensor([[tokenizer.bos_token_id]], dtype=torch.long).to(device) eos_token_id = torch.tensor([[tokenizer.eos_token_id]], dtype=torch.long).to(device) input_ids = torch.concat([bos_token_id, input_ids, eos_token_id], dim=1) with torch.no_grad(): outputs = model.generate( input_ids=input_ids, max_new_tokens=max_new_tokens, do_sample=True, top_p=top_p, temperature=temperature, repetition_penalty=repetition_penalty, eos_token_id=tokenizer.eos_token_id ) outputs = outputs.tolist()[0][len(input_ids[0]):] response = tokenizer.decode(outputs) response = response.strip().replace(tokenizer.eos_token, "").strip() print("Firefly：{}".format(response)) text = input('User：') if __name__ == '__main__': main() ``` 多轮对话: ```python from transformers import AutoModelForCausalLM, AutoTokenizer import torch def main(): model_name = 'YeungNLP/firefly-baichuan-13b' device = 'cuda' max_new_tokens = 500 # 每轮对话最多生成多少个token history_max_len = 1000 # 模型记忆的最大token长度 top_p = 0.9 temperature = 0.35 repetition_penalty = 1.0 # 加载模型 model = AutoModelForCausalLM.from_pretrained( model_name, trust_remote_code=True, low_cpu_mem_usage=True, torch_dtype=torch.float16, device_map='auto' ).to(device).eval() tokenizer = AutoTokenizer.from_pretrained( model_name, trust_remote_code=True, # llama不支持fast use_fast=False if model.config.model_type == 'llama' else True ) # QWenTokenizer比较特殊，pad_token_id、bos_token_id、eos_token_id均为None。eod_id对应的token为<|endoftext|> if tokenizer.__class__.__name__ == 'QWenTokenizer': tokenizer.pad_token_id = tokenizer.eod_id tokenizer.bos_token_id = tokenizer.eod_id tokenizer.eos_token_id = tokenizer.eod_id # 记录所有历史记录 if model.config.model_type != 'chatglm': history_token_ids = torch.tensor([[tokenizer.bos_token_id]], dtype=torch.long) else: history_token_ids = torch.tensor([[]], dtype=torch.long) # 开始对话 utterance_id = 0 # 记录当前是第几轮对话，为了契合chatglm的数据组织格式 user_input = input('User：') while True: utterance_id += 1 # chatglm使用官方的数据组织格式 if model.config.model_type == 'chatglm': user_input = '[Round {}]\n\n问：{}\n\n答：'.format(utterance_id, user_input) user_input_ids = tokenizer(user_input, return_tensors="pt", add_special_tokens=False).input_ids # firefly的数据组织格式 # 为了兼容qwen-7b，因为其对eos_token进行tokenize，无法得到对应的eos_token_id else: input_ids = tokenizer(user_input, return_tensors="pt", add_special_tokens=False).input_ids eos_token_id = torch.tensor([[tokenizer.eos_token_id]], dtype=torch.long) user_input_ids = torch.concat([input_ids, eos_token_id], dim=1) history_token_ids = torch.concat((history_token_ids, user_input_ids), dim=1) model_input_ids = history_token_ids[:, -history_max_len:].to(device) with torch.no_grad(): outputs = model.generate( input_ids=model_input_ids, max_new_tokens=max_new_tokens, do_sample=True, top_p=top_p, temperature=temperature, repetition_penalty=repetition_penalty, eos_token_id=tokenizer.eos_token_id ) model_input_ids_len = model_input_ids.size(1) response_ids = outputs[:, model_input_ids_len:] history_token_ids = torch.concat((history_token_ids, response_ids.cpu()), dim=1) response = tokenizer.batch_decode(response_ids) print("Firefly：" + response[0].strip().replace(tokenizer.eos_token, "")) user_input = input('User：') if __name__ == '__main__': main() ```