solaarphunk/turbospeak-correction-model

feat: v7 model — 3000 iters cosine LR, cleaned 8K data, 147/185 (79%)

feat: v5 model with versioned filename for auto-update

feat: v5 model — 8,253 training examples, 150/185 benchmark (81%)

Update model card with v12b benchmark results

Update to v12b model (87.5% correction accuracy on 40-case benchmark)

Upload qwen3-1.7b-correction-q4_k_m.gguf with huggingface_hub

initial commit