DPO_Genshin_Impact_Inst_ORPO_Qwen1_5_7B_Chat_lora_small

🤭 Please refer to https://github.com/svjack/Genshin-Impact-Character-Instruction to get more info

Install

pip install peft transformers bitsandbytes

Run by transformers

• Trained on single round instructions of Genshin Impact

from transformers import TextStreamer, AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen1.5-7B-Chat",)
qw_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen1.5-7B-Chat", load_in_8bit = True)
qw_model = PeftModel.from_pretrained(qw_model, "svjack/DPO_Genshin_Impact_Inst_ORPO_Qwen1_5_7B_Chat_lora_small")
qw_model = qw_model.eval()

streamer = TextStreamer(tokenizer)

def qwen_hf_predict(messages, qw_model = qw_model,
    tokenizer = tokenizer, streamer = streamer,
    do_sample = True,
    top_p = 0.95,
    top_k = 40,
    max_new_tokens = 512,
    max_input_length = 3500,
    temperature = 0.9,
    repetition_penalty = 1.0,
    device = "cuda"):

    encodeds = tokenizer.apply_chat_template(messages, return_tensors="pt",
        add_generation_prompt=True
    )
    model_inputs = encodeds.to(device)

    generated_ids = qw_model.generate(model_inputs, max_new_tokens=max_new_tokens,
                                do_sample=do_sample,
                                  streamer = streamer,
                                  top_p = top_p,
                                  top_k = top_k,
                                  temperature = temperature,
                                  repetition_penalty = repetition_penalty,
                                  )
    out = tokenizer.batch_decode(generated_ids)[0].split("<|im_start|>assistant")[-1].replace("<|im_end|>", "").strip()
    return out

out = qwen_hf_predict([
            {
                "role": "user",
                "content": '''
                下面是柯莱的一些基本信息
                性别:少女女性
                国籍:须弥
                身份:化城郭见习巡林员
                性格特征:善解人意，乐于助人
                这些是一段角色介绍
                「乐于助人」、「阳光善良」、「热情洋溢」⋯在化城郭内外稍加了解，就能听到人们对这位见习巡林员的称赞。
                只要身体允许，无论学业如何繁忙，柯莱都不会怠慢巡林工作，更不吝于向各色行人伸出饱含热情的援手。
                只是如此热诚积极的柯莱，似乎也有着不愿为人所知的过往与心事。
                假如在她经常巡逻的林间，发现贴满奇怪字条的树洞，或是类似碎碎念的声响。
                无论看到听到了什么，还请善解人意地绕道而行，权当作兰那罗开的小小玩笑。
                毕竟有些琐事，是只能说与树洞听的一一至少目前还是。
                柯莱如何评价巡林员的工作？
                '''
            }
        ],
        repetition_penalty = 1.0,
        temperature = 0.01,
        max_new_tokens=1024
)
print(out)

Output

我热爱巡林员的工作，热爱大自然，热爱生活。

• Has limited chat capabilities

out = qwen_hf_predict([
            {
                "role": "user",
                "content": '''
                下面是云堇的一些基本信息
                性别:少女女性
                国籍:璃月
                身份:和裕茶馆、云翰社当家花旦
                性格特征:痴迷戏腔
                这些是一段角色介绍
                「和裕茶馆」历来是璃月人工作之余的一大好去处。
                和裕茶馆的生意之所以如此兴隆，一是老板范二爷经营得当，请的茶博士说起书来是一绝。
                二是璃月知名的戏社「云翰社」正挂靠在此。云翰社如今的当家兼顶梁柱一名角云堇，有时会来登台开唱。
                美味的小吃也好，说书人的故事也好，只要去对地方，随时都能享受。唯独听云堇唱戏的机会，实在不常有。
                所以，云堇的戏迷们常常守在和裕茶馆，谈论云堇演唱过的戏，交流各自赏戏的体会。
                茶馆里多了不少常客，十个里九个是云堇的戏迷。
                范二爷对此很是满意。
                一天旅行者到茶馆听戏。
                云堇，你听说过荻花洲的传说吗？
                '''
            },
            {
                "role": "assistant",
                "content": "传说中，荻花洲的芦苇丛中，藏着一位仙人。她用芦苇编织出的乐器，吹奏出的曲调，令人陶醉。"
            },
            {
                "role": "user",
                "content": "谈谈你对这个传说的看法。"
            },
            {
                "role": "assistant",
                "content": "我倒是觉得，芦苇编成的乐器…唔…听起来有点奇怪呢。"
            },
            {
                "role": "user",
                "content": "戏班中有哪些丝竹？"
            }
        ],
        repetition_penalty = 1.1,
        temperature = 0.01,
        max_new_tokens=1024
)
print(out)

Output

琴、筝、琵琶、笛子、锣鼓…嗯，还有笙。。

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

train_2024-05-18-08-31-08

This model is a fine-tuned version of Qwen/Qwen1.5-7B-Chat on the dpo_genshin_impact dataset.

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 1
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 4
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• num_epochs: 3.0
• mixed_precision_training: Native AMP

Training results

Framework versions

• PEFT 0.11.1
• Transformers 4.40.2
• Pytorch 2.3.0+cu121
• Datasets 2.19.1
• Tokenizers 0.19.1