File size: 6,358 Bytes
905da01
 
 
 
 
 
5f8ec52
713360d
5f8ec52
713360d
5f8ec52
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1c3fedd
 
 
 
 
 
 
11eb5ea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1c3fedd
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
---
license: apache-2.0
language:
- en
base_model:
- codellama/CodeLlama-7b-hf
---
# [EMNLP 2025] TL-CodeLLaMA-2

TL-CodeLLaMA-2 is a model designed for tool use, built upon CodeLLaMA-7b. It is trained on 1,217 data samples using the *TL-Training* framework and demonstrates effective performance across a variety of tool use tasks. More information can be found in the paper "[TL-Training: A Task-Feature-Based Framework for Training Large Language Models in Tool Use](https://www.arxiv.org/abs/2412.15495)".

# Model Use

## Requirements

To use this model, please make sure to install transformers:
```bash
pip install transformers
```

## Data Orgnization

The data needs to be organized in the following format:

```json
[
    {
        "role": "System",
        "content": "Function:\ndef random_advice():\n    \"\"\"\n    Returns a random advice slip as a slip object.\n    \"\"\"\n\nFunction:\ndef advice_by_id(slip_id:str):\n    \"\"\"\n    If an advice slip is found with the corresponding {slip_id}, a slip object is returned.\n\n    Args:\n        slip_id (string): The unique ID of this advice slip.\n    \"\"\"\n\nFunction:\ndef search_advice(query:str):\n    \"\"\"\n    If an advice slip is found, containing the corresponding search term in {query}, an array of slip objects is returned inside a search object.\n\n    Args:\n        query (string): The search query provided.\n    \"\"\"\n\nFunction:\ndef ask_to_user(question:str):\n    \"\"\"\n    You can ask user for guidance when you think you need more information to handle the task, but you should use this tool as less as you can.\n\n    Args:\n        question (string): The question you want to ask to user.\n    \"\"\"\n\nFunction:\ndef finish(answer:str):\n    \"\"\"\n    Finish the task and give your answer.\n\n    Args:\n        answer (string): Your answer for the task.\n    \"\"\"\n\n"
    },
    {
        "role": "User",
        "content": "Could you give me some advice about 'love'?"
    },
    {
        "role": "Assistant",
        "content": "search_advice(query = 'love') "
    },
    {
        "role": "Output",
        "content": "..."
    }
]
```

## Chat Template

The chat template is:

```jinja
{% for message in messages %}{{message['role'] + ': ' + message['content']}}{% if loop.last %}{% if add_generation_prompt %}{{ '\nAssistant:' }}{% else %}{{ '</s>'}}{% endif %}{% else %}{{ '\n' }}{% endif %}{% endfor %}
```

## Inference

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_path = "Junjie-Ye/TL-CodeLLaMA-2"

data = [
    {
        "role": "System",
        "content": "Function:\ndef random_advice():\n    \"\"\"\n    Returns a random advice slip as a slip object.\n    \"\"\"\n\nFunction:\ndef advice_by_id(slip_id:str):\n    \"\"\"\n    If an advice slip is found with the corresponding {slip_id}, a slip object is returned.\n\n    Args:\n        slip_id (string): The unique ID of this advice slip.\n    \"\"\"\n\nFunction:\ndef search_advice(query:str):\n    \"\"\"\n    If an advice slip is found, containing the corresponding search term in {query}, an array of slip objects is returned inside a search object.\n\n    Args:\n        query (string): The search query provided.\n    \"\"\"\n\nFunction:\ndef ask_to_user(question:str):\n    \"\"\"\n    You can ask user for guidance when you think you need more information to handle the task, but you should use this tool as less as you can.\n\n    Args:\n        question (string): The question you want to ask to user.\n    \"\"\"\n\nFunction:\ndef finish(answer:str):\n    \"\"\"\n    Finish the task and give your answer.\n\n    Args:\n        answer (string): Your answer for the task.\n    \"\"\"\n\n"
    },
    {
        "role": "User",
        "content": "Could you give me some advice about 'love'?"
    }
]

chat_template = "{% for message in messages %}{{message['role'] + ': ' + message['content']}}{% if loop.last %}{% if add_generation_prompt %}{{ '\nAssistant:' }}{% else %}{{ '</s>'}}{% endif %}{% else %}{{ '\n' }}{% endif %}{% endfor %}"

model = AutoModelForCausalLM.from_pretrained(
    model_path,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True
).eval()

tokenizer = AutoTokenizer.from_pretrained(model_path,
                                            padding_side="left",
                                            trust_remote_code=True)
if tokenizer.pad_token_id is None:
    tokenizer.pad_token_id = tokenizer.eos_token_id

text = tokenizer.apply_chat_template(
        data,
        tokenize=False,
        chat_template=chat_template,
        add_generation_prompt=add_generation_prompt
    )
model_inputs = tokenizer(
    [text], return_tensors="pt", padding=True).to("cuda")

generated_ids = model.generate(
    max_new_tokens=1024,
    **model_inputs,
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)

print(response)
```

# Citation

If you find this model useful in your research, please cite:

```bibtex
@inproceedings{TL-Training,
  author       = {Junjie Ye and
                  Yilong Wu and
                  Sixian Li and
                  Yuming Yang and
                  Zhiheng Xi and
                  Tao Gui and
                  Qi Zhang and
                  Xuanjing Huang and
                  Peng Wang and
                  Zhongchao Shi and
                  Jianping Fan and
                  Zhengyin Du},
  editor       = {Christos Christodoulopoulos and
                  Tanmoy Chakraborty and
                  Carolyn Rose and
                  Violet Peng},
  title        = {TL-Training: {A} Task-Feature-Based Framework for Training Large Language
                  Models in Tool Use},
  booktitle    = {Findings of the Association for Computational Linguistics: {EMNLP}
                  2025, Suzhou, China, November 4-9, 2025},
  pages        = {239--258},
  publisher    = {Association for Computational Linguistics},
  year         = {2025},
  url          = {https://aclanthology.org/2025.findings-emnlp.15/},
  timestamp    = {Fri, 20 Feb 2026 08:07:46 +0100},
  biburl       = {https://dblp.org/rec/conf/emnlp/YeWLYXGZHWSFD25.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}
```