stevenvo780/autologic-slm-onnx

Autologic SLM - ONNX fp32

Fine-tuned Qwen2.5-0.5B model that translates natural language text into structured logic AST (JSON).

Model Details

• Base model: Qwen/Qwen2.5-0.5B
• Training: LoRA (r=16, alpha=32) on q/v/k/o_proj, bf16, 15 epochs
• Dataset: 395 verified samples, 10 logic patterns (modus ponens, syllogisms, etc.)
• Format: ONNX fp32 with KV cache (for Transformers.js)
• Size: ~2.4 GB
• RAM usage: ~2.2 GB at inference

Output Format

The model outputs JSON conforming to Autologic AST:

{
  "axioms": [
    {"name": "a1", "formulaJSON": {"type": "Connective", "operator": "IMPLIES",
      "left": {"type": "Atom", "id": "Llueve", "text": "llueve"},
      "right": {"type": "Atom", "id": "SueloMojado", "text": "suelo mojado"}}}
  ],
  "conclusions": [
    {"formulaJSON": {"type": "Atom", "id": "SueloMojado", "text": "suelo mojado"}}
  ]
}

Critical Notes

• Must use fp32: int8/uint8 quantization destroys instruction following on 0.5B models
• Must use exact training prompt format: ChatML with specific system prompt
• Exported with task=text-generation-with-past (includes KV cache)
• Tokenizer merges are strings (Transformers.js compatible)

Part of

Autologic - Automatic formalization of natural language to formal logic.