thewulf7/dharma-core-qwen2.5-7b-4bit

drift-mcp-core — Qwen2.5-7B-Instruct-4bit

Schema-free tool calling via LoRA. This adapter enables a Qwen2.5-7B-Instruct-4bit base model to call tools correctly without needing tool schemas in the prompt at inference time.

Part of the drift-engine project.

Eval Results (tools-v3)

The adapter matches or exceeds the baseline with zero schema tokens at inference.

Tools Covered

web_search, read_file, write_file, list_directory, terminal, fetch

Framework

This adapter is in MLX format (native Apple Silicon). It is NOT compatible with HuggingFace PEFT/transformers directly — use it with mlx-lm or drift serve.

# Install drift
pip install drift-engine

# Serve
drift catalog pull drift-mcp-core --model qwen2.5-7b-4bit
drift serve adapters/drift-mcp-core-qwen2.5-7b-4bit --base-model qwen2.5-7b-4bit

Training Details

• —Base model: mlx-community/Qwen2.5-7B-Instruct-4bit
• —Framework: mlx-lm LoRA
• —Dataset: 2,315 examples (2,083 train / 232 val)
• —1,855 tool-call positives + 460 clean negatives (19.8%)
• —LoRA rank: 8, alpha: 16, dropout: 0.05
• —Epochs: 3 (6,249 iters), batch size: 1, grad accum: 4
• —Learning rate: 1e-4
• —Val loss: 0.316

Key Finding

Negative example quality matters more than quantity. 33% ambiguous negatives → 0% no-call accuracy. 19.8% clean (unambiguous factual) negatives → 100% no-call accuracy.

Version

drift-mcp-core@1.0.0+qwen2.5-7b-instruct-4bit