codie-oracle
A 0.5B parameter routing model that maps natural language to structured symbolic decisions using the CODIE symbolic layer.
85.3% accuracy. 1GB VRAM. 16ms inference.
What it does
Given any natural language input, returns a deterministic JSON routing decision:
The 15 CODIE symbols
| Symbol | Name | Accuracy |
|---|---|---|
| ≟ | COND | 100% |
| □ | CONTAINER | 100% |
| ▶ | TRIGGER | 100% |
| ⟳ | LOOP | 95% |
| ± | STATE | 90% |
| ⊕ | MERGE | 90% |
| ⊙ | NODE | 90% |
| ⇐ | READ | 85% |
| ⇒ | WRITE | 85% |
| ⊗ | OP | 80% |
| ◇ | INPUT | 80% |
| ← | FLOW | 75% |
| ∑ | AGGREGATE | 75% |
| # | HASH_REF | 70% |
| ⊘ | NULL | 65% |
Overall: 85.3% (300-sample eval across all symbols)
Usage
Architecture
Training
- Base: Qwen2.5-0.5B-Instruct
- QLoRA: r=16, alpha=32, 4-bit NF4
- Corpus: 14,617 balanced pairs (harvested + synthetic + targeted patches)
- Hardware: RTX 3090 Ti, ~16 min per run
- 10 training rounds with per-symbol accuracy-guided corpus patching
GitHub
- Downloads last month
- 3
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct") model = PeftModel.from_pretrained(base_model, "tomlee3ddesign/Codie")