juanquivilla/sotto-cleanup-lfm25-350m

soup: corrected README with full prod metrics + soup explanation

soup_30: composite=89.45 — see model card for benchmark deltas vs v45

v55: corrected metrics with max_new=900 (sub_del15 3.7%, composite 89.32)

v55: add inference recommendations (rep_pen=1.05, max_new_tokens guidance)

v55: composite=88.95 — see model card for benchmark deltas vs v45

v51: composite=88.68 — see model card for benchmark deltas vs v45

v45: SFT+chained GRPO with ITN — 95.9% number accuracy, 97.0% filler-free, deletion behavior matches v36

v36: full-FT GRPO with substantive-deletion-aware reward — filler-free 96.9%, sub-del-15-long 0.64%

v23 R6 (4-stage SFT->GRPO->Stage2->GRPO): ROUGE-L 0.9537 (tied v22), Filler-Free 91.1% (beats v22 90.3%), paragraph rate 91.5% — definitive v23 model

v23 R5 (paragraph rows excluded from GRPO): ROUGE-L 0.9505, Filler-Free 91.0%, paragraph rate 89.5% — best v23 variant overall

v23 R4 (LR 5e-6): Filler-Free 91.0% (beats v22 90.3%), paragraph rate 89%, ROUGE-L 0.9499

v23+paragraphs: ROUGE-L 0.9506, Filler-Free 90.2%, paragraph rate 91.5% (0% in v22)

v22+GRPO-r32: ROUGE-L 0.954 val, 66% exact, 91% filler-free — new production model

v22+GRPO: ROUGE-L 0.953 val set, 91% filler-free — GRPO works on proper benchmark

v22-lr3: ROUGE-L 0.948 on val set — cleaned data + LR 3e-5

v18: ROUGE-L 0.968, 72% exact — AdamW beta2=0.95 breakthrough

v17: ROUGE-L 0.966, 71% exact — preserve phrases + redundant tail removal

v16: ROUGE-L 0.962, preserve-phrase data (crutch_words 0.987)

Update model card: v15 detailed card with examples, benchmarks, and links

v15: ROUGE-L 0.960, 70% exact match — LR 2.5e-5 breakthrough

v7: combined dataset LR 2e-5, ROUGE-L 0.943, 62% exact

v5: LR 2e-5 + Stage-2, ROUGE-L 0.942, 60% exact — new all-time record

v4: 116K data + long transcripts, ROUGE-L 0.927, handles 1000+ word inputs

Upload README.md with huggingface_hub

Stage-2: ROUGE-L 0.931, 90% zero-filler — new best

Full FT v2: ROUGE-L 0.930, 55% exact — best model

Full FT + GRPO: ROUGE-L 0.916 — new record, +2.5 over 2B

Full FT: ROUGE-L 0.907 — new record, +1.6 over prompted 2B

GRPO R2: ROUGE-L 0.892 — exceeds prompted 2B

GRPO model: ROUGE-L 0.891 — matches prompted 2B

Upload folder using huggingface_hub

Upload README.md with huggingface_hub

initial commit