qwen25-3b-openclaw

A Qwen2.5-3B-Instruct model fine-tuned for exceptional tool/function calling ability, purpose-built as the local agent model for OpenClaw / LocalClaw served via vLLM.

Model Summary

Property	Value
Base model	Qwen/Qwen2.5-3B-Instruct
Fine-tuning method	LoRA (rank=16, alpha=32, all 32 layers)
Training framework	mlx-lm on Apple M4 Max
Training data	~57k tool-call examples (hermes-function-calling-v1 + glaive-function-calling-v2)
Training steps	600
Tool format	<tool_call> / </tool_call> (Qwen/Hermes convention)
Serving target	vLLM with --enable-auto-tool-choice --tool-call-parser hermes

Evaluation

Evaluated on a held-out set of 50 tool-calling examples from the training distribution.

Metric	Score
tool_score (composite)	0.989
name_accuracy	1.000
arg_f1	0.983
parse_rate	0.980
val_loss	0.010

• name_accuracy: fraction of examples where the correct function name was called
• arg_f1: token-level F1 between predicted and ground-truth arguments
• parse_rate: fraction of outputs that contained a valid, parseable <tool_call> block
• tool_score: name_accuracy × 0.4 + arg_f1 × 0.4 + parse_rate × 0.2

What It's Good For

• OpenClaw / LocalClaw agent — drop-in local model for the tool-calling tier; handles calendars, email, web search, browser control, and custom skill tools
• Any OpenAI-compatible tool-use pipeline — responds to the standard tools parameter and produces structured function calls
• Offline / privacy-first deployments — 3B parameters runs fast on Apple Silicon or a modest GPU; no cloud dependency
• Multi-tool selection — trained on examples with multiple available tools; reliably selects the right one
• Argument extraction — near-perfect extraction of typed arguments from natural-language queries

What It's Not Good For

• Long multi-turn reasoning chains (consider a larger model for orchestration)
• Tasks requiring no tools — the model is biased toward calling tools when they're available
• Languages other than English (training data is English-only)

Tool Call Format

The model outputs tool calls using the Hermes <tool_call> convention:

<tool_call>
{"name": "function_name", "arguments": {"arg1": "value1", "arg2": 42}}
</tool_call>

Multiple parallel calls are output sequentially, each in its own block.

The system prompt should include available tools in <tools></tools> XML tags:

You are a function calling AI model. You are provided with function signatures
within <tools> </tools> XML tags. You may call one or more functions to assist
with the user query.

<tools>
{"type": "function", "function": {"name": "get_weather", "description": "...", "parameters": {...}}}
</tools>

For each function call return a json object within <tool_call> </tool_call> tags.

Quick Start with vLLM

pip install vllm

vllm serve sunkencity/qwen25-3b-openclaw \
  --enable-auto-tool-choice \
  --tool-call-parser hermes \
  --port 8000

This exposes an OpenAI-compatible API at http://localhost:8000/v1 with structured tool_calls in responses.

Integration with OpenClaw / LocalClaw

LocalClaw auto-discovers models from a running vLLM server. After starting the server above, add to ~/.localclaw/openclaw.local.json:

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "vllm/sunkencity/qwen25-3b-openclaw"
      }
    }
  }
}

LocalClaw will route tool-calling tasks to this model automatically via the three-tier routing system.

Python Usage (mlx-lm)

from mlx_lm import load, generate

model, tokenizer = load("sunkencity/qwen25-3b-openclaw")

messages = [
    {
        "role": "system",
        "content": (
            "You are a function calling AI model.\n\n"
            "<tools>\n"
            '{"type": "function", "function": {"name": "get_weather", '
            '"description": "Get current weather", "parameters": {"type": "object", '
            '"properties": {"location": {"type": "string"}}, "required": ["location"]}}}\n'
            "</tools>\n\n"
            "For each function call return a json object within <tool_call> </tool_call> tags."
        ),
    },
    {"role": "user", "content": "What is the weather in San Francisco?"},
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
response = generate(model, tokenizer, prompt=prompt, max_tokens=256, verbose=False)
print(response)
# <tool_call>
# {"name": "get_weather", "arguments": {"location": "San Francisco"}}
# </tool_call>

Training Details

Datasets:

• NousResearch/hermes-function-calling-v1 — 1,883 examples, already in <tool_call> format
• glaiveai/glaive-function-calling-v2 — 55,000 examples converted from <functioncall> format

Preprocessing:

• Conversations truncated to the first tool-call turn (system + user + assistant) to ensure tool call output is always within the sequence budget
• <functioncall> → <tool_call> normalization including balanced-brace extraction and single-quoted argument repair
• 90/5/5 train/valid/eval split

LoRA config:

fine_tune_type: lora
num_layers: 32          # all transformer layers
lora_parameters:
  rank: 16
  alpha: 32
  dropout: 0.0
  scale: 10.0
optimizer: adamw
learning_rate: 2e-4
batch_size: 4
iters: 600
max_seq_length: 2048
mask_prompt: true       # loss only on assistant turns
grad_checkpoint: true

Hardware: Apple M4 Max (128 GB unified memory), ~34 minutes wall-clock

Training Code

Training was performed using the tool-tuner framework — an autoresearch-inspired autonomous LoRA experiment loop built with mlx-lm.

License

Apache 2.0 — same as base model.