qwen2.5-0.5b-funccall
A fine-tuned Qwen2.5-0.5B-Instruct that takes a user query plus a set of available tool/function schemas and outputs the correct function call(s) as clean, parseable JSON — no prose, no markdown fences. Trained as a cheap, accurate "router" model: given a natural-language request and a list of tools, it picks the right tool and fills in arguments correctly, so you don't need to call a much larger model on every turn.
Model details
- Base model:
unsloth/Qwen2.5-0.5B-Instruct - Method: LoRA fine-tuning via Unsloth, merged into the base weights
- Training data:
Salesforce/xlam-function-calling-60k— 60k function-calling examples, each verified through format checking, real function execution, and semantic verification - Task framing: system message lists available tools as JSON → user message is the natural-language query → assistant response is a JSON array of
{"name": ..., "arguments": ...}objects, and only that
Intended use
Drop-in tool/function router for agent loops, CLI dispatchers, or any system that needs to map a user request to a structured function call without paying for a large general-purpose model on every request.
Why this model
Salesforce's own xLAM-1b-fc-r already showed that a sub-2B model can place competitively on the Berkeley Function-Calling Leaderboard (BFCL), outperforming several much larger general-purpose models. This model explores the same idea at an even smaller scale (0.5B), using Unsloth for fast LoRA fine-tuning.
Evaluation status
Not yet evaluated. Internal exact-match scoring (function name + argument match on a held-out split of the training data) is in progress. The model has not yet been benchmarked against Salesforce/xLAM-1b-fc-r or against larger zero-shot baselines (e.g. Qwen2.5-7B-Instruct) on the real BFCL harness. Numbers below will be filled in once that's run — treat any claims of "matching" or "beating" larger models as not yet verified until this section is updated.
| Model | BFCL category | Accuracy |
|---|---|---|
nakue/qwen2.5-0.5b-funccall (this model) | simple, multiple | pending |
Qwen2.5-7B-Instruct (zero-shot) | simple, multiple | pending |
Salesforce/xLAM-1b-fc-r | simple, multiple | pending |
How to use
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch, json
def build_xlam_system_prompt(tools_xlam_format):
"""
tools_xlam_format: a list of tool dicts already in xLAM-native shape, e.g.
[
{
"name": "get_weather",
"description": "Get current weather for a location",
"parameters": {
"location": {
"description": "The city to get weather for",
"type": "str"
}
}
}
]
"""
return (
"You are a function-calling assistant. Given a user query and a list of "
"available tools, respond with ONLY a JSON array of the function call(s) "
"needed to fulfill the query. Each item must have 'name' and 'arguments' "
"keys. Do not include any explanation, markdown formatting, or text other "
f"than the raw JSON array.\n\nAvailable tools:\n{json.dumps(tools_xlam_format, indent=2)}"
)
# Example usage with your weather tool, written directly in xLAM format:
tools = [
{
"name": "get_weather",
"description": "Get current weather for a location",
"parameters": {
"location": {
"description": "The city to get weather for",
"type": "str"
}
}
}
]
system_msg = build_xlam_system_prompt(tools)
messages = [
{"role": "system", "content": system_msg},
{"role": "user", "content": "What's the weather like in Harare right now?"},
]
inputs = tokenizer.apply_chat_template(
messages, add_generation_prompt=True, return_tensors="pt", return_dict=True
).to(model.device)
out = model.generate(
**inputs,
max_new_tokens=256,
do_sample=False,
pad_token_id=tokenizer.eos_token_id,
)
response = tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(response)
Limitations
- Trained only on
simpleandmultiple-style single-turn function calling. Not trained or tested on multi-turn conversations, parallel calls, or queries with no matching tool (irrelevance detection). - Output is sensitive to tool-schema formatting; large or unusual schemas outside the training distribution may degrade reliability.
- Evaluation against published baselines is pending — see above.
Training details
Fine-tuned with Unsloth's LoRA implementation (r=16, lora_alpha=16, targeting attention and MLP projection layers), 2 epochs, cosine LR schedule, on a held-out-respecting split of xlam-function-calling-60k (500 examples reserved for test, 300 for validation, remainder for training).
Citation
If you use this model, please also cite the underlying dataset:
@misc{xlam,
title={xLAM: A Family of Large Action Models to Empower AI Agent Systems},
author={Salesforce AI Research},
year={2024}
}