debashis/llama-1b-tool-router-grpo

LLaMA 1B Tool Router - GRPO Trained

A 1B parameter model trained with pure PyTorch GRPO for fast, accurate tool selection in agentic AI systems.

Results

Metric	Value
Validation Accuracy	87.94%
Parameters	1B
Training	GRPO only (no SFT)
Dataset	Salesforce xLAM 60k

Why This Model?

	GPT-4 / Claude API	This Model
Latency	2-5 seconds	100-500ms
Cost/call	$0.01-0.03	~$0.0001
Privacy	Cloud	Local
Offline	❌	✅

100x cheaper. 10x faster. Runs anywhere.

Quick Start

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "debashis/llama-1b-tool-router-grpo",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("debashis/llama-1b-tool-router-grpo")

def call_tool(query: str, tools: list) -> str:
    tools_text = ""
    for i, tool in enumerate(tools, 1):
        tools_text += f"{i}. {tool['name']}\n"
        tools_text += f"   Description: {tool['description']}\n"
        tools_text += f"   Parameters: {', '.join(tool['parameters'])}\n"
    
    prompt = f"""<|begin_of_text|><|start_header_id|>system<|end_header_id|>

You are a function calling assistant. Based on the user's query, call the appropriate function with correct arguments.

Available functions:
{tools_text}
Respond with a JSON object: {{"name": "function_name", "arguments": {{"param1": "value1"}}}}

Only output JSON.<|eot_id|><|start_header_id|>user<|end_header_id|}

{query}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

"""
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    outputs = model.generate(**inputs, max_new_tokens=150, temperature=0.1, do_sample=False)
    return tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)

Example

tools = [
    {"name": "search_flights", "description": "Search for flights", 
     "parameters": ["origin", "destination", "date"]},
    {"name": "book_hotel", "description": "Book a hotel room", 
     "parameters": ["city", "checkin", "checkout"]},
]

result = call_tool("Find flights from NYC to London on Dec 25th", tools)
# Output: {"name": "search_flights", "arguments": {"origin": "NYC", "destination": "London", "date": "Dec 25th"}}

Training Details

Parameter	Value
Base model	LLaMA 3.2 1B Instruct
Algorithm	GRPO (Pure PyTorch)
Epochs	3
Learning rate	1e-6
KL coefficient	0.1
Temperature	0.7 (train) / 0.1 (inference)

Training Curves

Limitations

• May struggle with 10+ similar tools
• Complex nested objects may lose detail
• Domain-specific tools benefit from additional fine-tuning

Links

• Training Code: GitHub
• Dataset: Salesforce/xlam-function-calling-60k

Citation

@misc{llama-tool-router-grpo-2025,
  author = {Debashis Ghosh},
  title = {LLaMA 1B Tool Router: GRPO-Trained for Agentic Systems},
  year = {2026},
  publisher = {Hugging Face}
}