jatshi/trusted-full-duplex-agent-data

Trusted Full-Duplex Speech Agent — Training Data & Evidence

Companion dataset for model: jatshi/trusted-full-duplex-agent and GitHub: Jatshi/trusted-full-duplex-agent.

2.0 evidence update

The 2.0 release adds the evidence needed to reproduce the deployed runtime rather than only the offline research pipeline: real-time latency telemetry, unified guardrail voice samples, the exact turn-taking MLP reports, TrustGate/ASR runtime configuration, and integrity manifests. The base MiniCPM-o, Qwen and Faster-Whisper checkpoints remain referenced upstream and are not redistributed.

See the versioned v2.0/ directory in the companion model repository for the GRPO LoRA adapter and project-owned runtime assets.

Contents

GRPO training data

File	Rows	Task
rl_streams/bargein_context.jsonl	60	barge-in context recall: each row = user request → truncated mid-sentence assistant turn (dangling, no end token) → user barge-in → probe instruction ("打断之前你说到哪了？"). 5 topics × 3 truncation points × 4 barge-in phrasings. interrupted_content field is the ground truth for the bigram-Jaccard context_recall metric.
rl_streams/rl_streams.jsonl	300	turn-taking / guardrail alignment samples (execute/clarify/stop families)

Turn-taking decision training

File	Description
turntaking/frames.npz	200ms-frame features + hold/take/backchannel labels (synthesized pause patterns: filler vs. final pauses, energy decay, pre-pause drop). Trains the external turn-taking MLP (96.5% frame accuracy, 0% false speaking rate).

User turn wavs (16kHz, Windows SAPI Huihui)

user_turns/*.wav — 5 guardrail scenarios + 3 barge-in turns (open / stop / probe) used for the real-base duplex session and the barge-in A/B attribution experiment (--bargein-delay-ms 500 → 5000).

Real-base evidence (MiniCPM-o 4.5 9B, AutoDL RTX 4080 SUPER 32GB)

duplex_session/ — five-scenario guardrail session audio, barge-in report JSONs (immediate vs. delayed interruption) and turn wavs backing the A/B attribution finding: probe bigram overlap 0.0 → 0.089 when the bot played 6s instead of 1s before being interrupted; recall shows recency bias + hallucination of un-played content.

Reproduce

git clone https://github.com/Jatshi/trusted-full-duplex-agent
cd trusted-full-duplex-agent
python scripts/33_prep_bargein_rl_data.py        # regenerates bargein_context.jsonl
python scripts/31_prep_rl_data.py                # regenerates rl_streams.jsonl
python scripts/60_turntaking_train.py            # retrains the MLP from frames.npz

Seed = 42 everywhere (see configs/rl.yaml).