README.md

---
license: apache-2.0
language:
- en
library_name: transformers
pipeline_tag: text-generation
tags:
- Hare
datasets:
- cerebras/SlimPajama-627B
- HuggingFaceTB/cosmopedia
arxiv: 2406.11410v1
---

<a id="english"></a>

<p align="center">
<img width="400px" alt="Lite-AI" src="./logo.jpg">
</p>

</div>


## Hare-1.1B-Tool
- Hare-1.1B-Tool is a fine-tuned version of [Hare-1.1B-base](https://huggingface.co/LiteAI/Hare-1.1B-base), designed to enable the invocation of Android system APIs and tool orchestration in composite scenarios on mobile devices. For a detailed introduction, please refer to [Hare-1.1B-base](https://huggingface.co/LiteAI/Hare-1.1B-base).
- Hare-1.1B-Tool是由[Hare-1.1B-base](https://huggingface.co/LiteAI/Hare-1.1B-base)微调而来，用于在手机端实现安卓系统API调用和组合场景下的工具调用。详细介绍请看[Hare-1.1B-base](https://huggingface.co/LiteAI/Hare-1.1B-base)。

## 模型使用
```python
import time
from transformers import GenerationConfig
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig, StoppingCriteria, StoppingCriteriaList
import os
import json
import logging
import torch

log_path = 'your_log_path'
logging.basicConfig(filename=log_path, level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')

logging.info('This is a log message.')

model_path = "/LiteAI/Hare-1.1B-Tool"
tokenizer = AutoTokenizer.from_pretrained(save_path)
model = AutoModelForCausalLM.from_pretrained(save_path)

class MyStoppingCriteria(StoppingCriteria):
    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
        keyword = tokenizer.decode(input_ids[0][-1])
        return keyword in ["<api_end>"]


def chat(
    messages,
    model,
    tokenizer,
    generate_config=None,
    max_length=512,
    max_new_tokens=256,
):
    if generate_config is None:
        generate_config = GenerationConfig(
            do_sample=False,
            max_length=max_length,
            max_new_tokens=max_new_tokens,
            eos_token_id=32001,
        )
    
    if messages[0]["role"] == "system":
        system = messages[0]["content"]
        messages = messages[0:]
    else:
        system = "You are a helpful assistant."
    
    n_token = max_length
    system = "<round_start>system\n{}<round_end>\n".format(system)
    system_token = tokenizer.encode(system, add_special_tokens=False)
    n_token -= len(system_token)

    query = messages[-1]["content"]
    query = "<round_start>user\n{}<round_end>\n<round_start>assistant\n".format(query)
    query_token = tokenizer.encode(query, add_special_tokens=False)
    n_token -= len(query_token)
    
    messages = messages[:-1]
    conversations = []
    for ids in range(len(messages)-1, 0, -2):
        user = messages[ids - 1]["content"]
        assistant = messages[ids]["content"]
        
        round = "<round_start>user\n{}<round_end>\n<round_start>assistant\n{}<round_end>\n".format(user, assistant)
        round_token = tokenizer.encode(round, add_special_tokens=False)

        if n_token - len(round_token) > 0:
            conversations = [round] + conversations
        else:
            break

    prompt = system + "".join(conversations) + query
    prompt_token = tokenizer(prompt, add_special_tokens=False, return_tensors="pt")
    prompt_token.to(model.device)


    response = model.generate(
        generation_config=generate_config,
        **prompt_token
    )

    output_tokens = response[0].cpu().numpy()[prompt_token.input_ids.size()[1]:]
    output_string = tokenizer.decode(output_tokens, skip_special_tokens=True).replace("<round_end>", "")
    return output_string, prompt

# ======================
#       main
# ======================

test_query_path = "/home/sxw/sft_exper/dataset/query_to_test"
for file in os.listdir(test_query_path):
    file_pth = os.path.join(test_query_path, file)
    print(file_pth)
    logging.info(file_pth)
    with open(file_pth, 'r') as f:
        for line in f:
            data = json.loads(line)
            # print(data)
            query = data["human"]
            messages = [
                {"role": "system", "content": "Below is the query from the users, you need make full sense of user's intention based on the content of the sentence, then call the correct function and generate the parameters of the calling function."},
                {"role": "user", "content": query}
            ]
            response, input_prompt = chat(messages=messages, model=model, tokenizer=tokenizer)

            logging.info(query)
            logging.info(data["assistant"])
            logging.info(response)
```