TealKit MCP LoRA Adapters

LoRA fine-tune of Qwen/Qwen2.5-3B-Instruct for structured MCP tool-call generation.

Trained for use with the TealKit agentic AI app — works in TealKit's server mode (remote inference via Ollama) and in local mode on any device where the fused GGUF fits in memory.

How to use: Download the adapters/ folder, fuse + convert to GGUF locally with bash scripts_training/train_mcp.sh --skip-train (see TealKit repo), then run via Ollama.

Training

• Dataset: custom MCP tool-call JSONL (125 train / validation examples)
• Method: QLoRA 4-bit NF4, LoRA r=16 α=32, target modules q_proj / v_proj
• Epochs: 1
• Batch: 2 × 2 grad-accum = 4 effective
• Hardware: Google Colab T4 (16 GB)
• Use case: MCP tool-calling for TealKit server mode and local on-device inference

Usage

from peft import PeftModel
from transformers import AutoTokenizer, AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-3B-Instruct")
model = PeftModel.from_pretrained(base, "lschaffer/tealkit", subfolder="adapters")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-3B-Instruct")

Training Hyperparameters

• r=16, α=32, LoRA dropout=0.05
• target_modules: q_proj, v_proj
• Learning rate: 2e-4, cosine scheduler
• Batch size: 2, gradient accumulation: 2
• Max seq length: 1024