FSI-ECHO-Gold

2.6M parameter code generation model — trains in ~1 hour on CPU, runs on any device.

Model Details

• Architecture: FSI_ECHO (Morph Embedding + Nanobot Swarm + Self-Attention + Assembly Blocks)
• Parameters: 2,621,578
• Vocab Size: 4,096
• Context Length: 2,048
• Training Data: 280 hand-verified Python function examples (80/20 train/val split)
• Training: Gold-standard curriculum — simple patterns first → multi-line → classes
• Export: GGUF (13MB f32), suitable for quantization

Files

File	Description
fsi_echo.py	Model architecture + tokenizer
gold_train.py	Gold-standard training pipeline with resume, early stopping, eval, export
exec_eval.py	Execution-based evaluation harness
gold_best.pt	Best checkpoint (val_loss=0.0836)
gold_final.pt	Final checkpoint (early stop at step 5600)
gold_fsi_echo.f32.gguf	GGUF export (13MB)
requirements.txt	PyTorch dependency

Quick Start

import sys; sys.path.insert(0, '.')
from fsi_echo import FSIEchoModel, CodeTokenizer
import torch

model = FSIEchoModel()
tok = CodeTokenizer()
ckpt = torch.load('gold_best.pt', map_location='cpu', weights_only=True)
model.load_state_dict(ckpt['model'])

# Generate code
result = model.generate(tok, "def factorial", max_tokens=80, temperature=0.1)
print(result['generated'])

Training

python3 gold_train.py

Resumes from latest checkpoint automatically. Trains to convergence with early stopping.

Accuracy

96.3% syntactic accuracy | See exec_eval.py for execution-based evaluation.

License

MIT