hf-wzx1205/PDO_S2TT-reproduction

PDO_S2TT — independent reproduction artifacts

This repository archives one fresh end-to-end reproduction run of the public PDO_S2TT reinforcement-learning code: training logs, a resumable training state (including optimizer state), the resulting inference checkpoint, and complete five-direction FLEURS TEST outputs.

[!IMPORTANT] These are reproduction artifacts, not the checkpoint used for the paper tables. The official paper checkpoint remains in `hf-wzx1205/PDO_S2TT`.

Run receipt

The inference checkpoint SHA-256 is bd6bc20f5a8dddd33991822bc17015b985932193abb0f7496d4e5a4df2081129. The resumable training-state SHA-256 is c3e94d3be46593cbf32bf66a5f5e55fd481484020cd1217ba3dfc4a466ae0e79.

Files

training/
├── pdo_s2tt_reproduced.pt   # inference-only checkpoint
├── training_state.pt        # policy + AdamW optimizer + counters
├── config.json              # exact public hyperparameters
├── progress.json            # final counters and fingerprint
├── complete.json            # terminal receipt
└── train.log                # complete sanitized training log

evaluation/
├── metrics/                 # complete metrics for Zh/De/Es/Ja/Fr
├── predictions/             # all 647 trajectories per direction
├── logs/                    # successful evaluator logs
├── macro.json               # five-direction aggregation
├── test.log                 # compact human-readable test receipt
└── audit.json               # integrity audit

training_state.pt contains the policy state, AdamW optimizer state, completed rollout-round counter, and completed-update counter. It is provided separately from the lighter inference checkpoint so the run can be inspected or resumed without confusing it with the paper model.

Download

hf download hf-wzx1205/PDO_S2TT-reproduction \
  --local-dir reproduction-artifacts

To use only the independently reproduced inference checkpoint:

hf download hf-wzx1205/PDO_S2TT-reproduction \
  training/pdo_s2tt_reproduced.pt \
  --local-dir reproduction-artifacts

Resume or inspect the training state

Clone the code repository and prepare the released five-direction FLEURS TRAIN manifest as described in its training guide. Then place the downloaded artifacts in the repository and pass the state explicitly:

torchrun --standalone --nproc-per-node=4 scripts/train_pdo.py \
  --manifest data/fleurs/train/train.jsonl \
  --sft-checkpoint checkpoints/pdo_s2tt_sft.pt \
  --output results/reproduction \
  --resume reproduction-artifacts/training/training_state.pt

The archived state is already at the published 407-round stopping point. Resuming with the unchanged recipe therefore verifies/restores terminal state; extending the schedule requires an explicit code/config change and is no longer the frozen reproduction protocol.

FLEURS TEST results

Small deviations from the paper checkpoint are expected because PDO uses stochastic on-policy sampling. This run independently recovered the same quality/latency operating point with full output coverage.

The archived trajectories were rescored on 2026-09-23 with evaluator commit `7636709`. This corrects Japanese kana and accented-Latin tokenization in LAAL-CU; quality, FTL, FRD, RTF, and revision metrics are unchanged.

Provenance and privacy

• —Predictions preserve every visible-draft event and measured compute time.
• —Local audio references were rewritten as portable data/fleurs/audio/test/*.wav paths; audio itself is not redistributed.
• —Logs were sanitized only to remove machine-specific absolute paths. Training and evaluation messages, counters, metrics, warnings, and failures/success states were otherwise preserved.
• —No access token, private credential, source audio, or paper checkpoint is included.

License

Code is released under Apache-2.0 in the main repository. Model use remains subject to the licenses and terms of the upstream base model and datasets.