stillerman/fdt-disfluency-distilbert-66m-v5

Disfluency deletion tagger for live speech transcripts: tags every whitespace word KEEP / DELETE / KEEP_STRIP_COMMA / KEEP_CAPITALIZE, then a ~15-line reconstruction turns tags into cleaned text. Deletion-only by construction — it cannot rephrase, hallucinate, or alter names and numbers.

• Architecture: distilbert-base-cased (65.2M params), v5 = v4 core-scope policy plus the 'like' round: quotative 'like' ('she was like do not touch that') and similative 'like' ('and like items') are KEEP schemas, paired against comma-marked filler 'like' after a be-verb ('it was, like, cold'). Steve Bench v2 as int8 ONNX at threshold 0.5: 592/650 (v4: 575); on the core-policy bench 611/650 (v4: 593). Quotative 'like' 69/72 (v4: 58/72)
• Val metrics: exact-match 0.9745, DELETE-F1 0.9916
• Training data: synthetic disfluency injection over conversational corpora — see stillerman/fdt-disfluency-synthetic
• onnx/model_quantized.onnx (int8) is ready for transformers.js (device: "webgpu", dtype: "q8"); runs at ~10–50 ms per utterance in-browser.
• ⚠️ Trained partly on DailyDialog (CC BY-NC-SA): treat as research artifact, not for commercial deployment as-is.

Trained on a DGX Spark as part of the FluencyAI digital-twin project.