padmanabh/mistralclaw-7b-tool-calling

MistralClaw - Mistral 7B Fine-tuned for MCP Tool Calling

Fine-tuned Mistral-7B-Instruct-v0.2 for reliable tool/function calling, optimized for OpenClaw MCP orchestration.

Model Details

• Base Model: mistralai/Mistral-7B-Instruct-v0.2
• Training: LoRA (rank 64, alpha 128) via Together AI
• Training Data: 12,054 examples from xlam, glaive, hermes-fc, OpenHermes
• Training Loss: 0.4219 (3 epochs)

Quick Start with Ollama

# 1. Pull base model
ollama pull mistral:7b-instruct-v0.2-q4_K_M

# 2. Download the LoRA adapter GGUF
# (download mistralclaw-lora.gguf from this repo)

# 3. Create Modelfile
cat > Modelfile << 'EOF'
FROM mistral:7b-instruct-v0.2-q4_K_M
ADAPTER ./mistralclaw-lora.gguf

PARAMETER temperature 0.1
PARAMETER top_p 0.9
PARAMETER num_predict 1024

SYSTEM """You are OpenClaw, a helpful AI assistant with access to tools. Use tools when needed by responding with JSON tool calls. For knowledge questions, answer directly."""
EOF

# 4. Create and run
ollama create mistralclaw -f Modelfile
ollama run mistralclaw

Files

File	Description	Size
mistralclaw-lora.gguf	LoRA adapter in GGUF format (for Ollama)	~320MB
adapter_model.safetensors	LoRA adapter weights (SafeTensors)	~640MB
adapter_config.json	LoRA configuration	1KB
Modelfile	Ollama Modelfile	1KB

Training Details

Trained on Together AI's A100 GPUs using LoRA fine-tuning on a mix of function-calling datasets:

• Salesforce xlam-function-calling-60k
• Glaive function-calling-v2
• NousResearch hermes-function-calling-v1
• OpenHermes multi-turn conversations

W&B Dashboard: https://wandb.ai/padmanabhg-freelance/together