gcode-specialist-3b

LoRA fine-tune of bigcode/starcoder2-3b on 709 prompt to G-code pairs (CNC milling / drilling operations across RS274/NGC, Fanuc, Haas, Siemens, Grbl dialects). Exported to Q4_K_M GGUF for in-process inference via llama-cpp-python on the g-gen HF Space.

Hyperparameters

• LoRA rank: 16, alpha: 32, dropout: 0.05
• target_modules: q_proj, k_proj, v_proj, o_proj, c_fc, c_proj
• max_seq_len: 2048, batch: 1 x 16 grad-accum
• epochs: 3, lr: 0.0002 (cosine, 10 warmup steps)
• 4-bit NF4 base via bitsandbytes, fp16 compute

Prompt format

<|system|>
You are a CNC G-code expert. Given a natural language description of a machining operation, output only valid G-code with no explanation, no markdown, and no prose. Each line must be a valid G-code block.
<|user|>
<your machining description>
<|assistant|>
<G-code>

Inference (llama-cpp-python)

from llama_cpp import Llama
llm = Llama(model_path="gcode-specialist-q4_k_m.gguf", n_ctx=2048)
out = llm.create_completion(
    "<|system|>\n" + SYSTEM_PROMPT + "\n<|user|>\nDrill 4 holes...\n<|assistant|>\n",
    max_tokens=512, stop=["M30\n", "<|user|>"],
)
print(out["choices"][0]["text"])