This repo contains a TRL SFT script to reproduce/adapt ConvFill: Model Collaboration for Responsive Conversational Voice Agents (arXiv:2511.07397) with Qwen/Qwen3.5-4B instead of SmolLM2-360M-Instruct.

Important: the paper's training corpus/code are not public yet. The script expects the same raw JSONL schema described in Section 4.1: one conversation per line, with turns containing user, responder (list of conversational phrases), and responder_thoughts (list of knowledge chunks or <|sil|>).

Training recipe implemented from the paper: 5 epochs, lr 1e-5, 500 warmup steps, cosine scheduler, added <|sil|> token, per-phrase prompt/completion expansion, completion/assistant-only SFT loss, Trackio logging, push to Hub.

Script: convfill_qwen35_repro.py.