| --- |
| pretty_name: rl final |
| language: |
| - en |
| - zh |
| tags: |
| - reinforcement-learning |
| - multimodal |
| - video |
| - image |
| --- |
| |
| # rl final |
|
|
| Mixed image/video RL release based on `mixed_rl460_rulerv2_20260915`, with the full training inputs and five complete launch commands. |
|
|
| The repository is publicly visible. Downloads do not require an access request. |
|
|
|
|
| ## Required code update — 2026-09-18 |
|
|
| The original `core.tar` still contains the September 15 source. **Apply the cumulative [context guard + video recovery fix](fixes/README.md) after extracting the archives and before starting a new training process.** The [TRANSFER.md](TRANSFER.md) procedure now includes this required step. Existing customized checkouts are checked before patching; unknown local changes are refused rather than overwritten. |
|
|
| The update preserves the original eight archives, environment, data, model, launch parameters and ablation settings. It includes the complete reference source, both baseline patches, the separate eval patch, trajectory evidence and CPU tests. The six changed production files were applied to source extracted from this repository's pinned `core.tar`; all **134 targeted CPU checks passed**. The archived SGLang validator source also matches the test fixture's source SHA. No GPU training or running process was changed. First-turn oversized inputs with zero generated tokens remain an explicit unsupported boundary. |
|
|
|
|
| - Eight original migration TAR archives: environment, code and image/video validation media. See [TRANSFER.md](TRANSFER.md). |
| - Image RL training data: 90 visual-probe shards and 277 base shards. |
| - Video RL training data: 27 Parquet shards containing 5,384 questions, plus exactly the 1,445 original MP4s selected by the manifest. |
| - [Five complete launch commands](commands/rl_five_commands_20260915/README.md), with `SAVE_FREQ=10` and `TEST_FREQ=10` in all five scripts. |
|
|
| See [DATA_LAYOUT.md](DATA_LAYOUT.md) for the dataset paths and video metadata conversion command. [RELEASE_FILES.json](RELEASE_FILES.json) records file sizes, SHA-256 hashes, and pinned upstream revisions for the supplemental files. |
|
|
| The original migration archives exclude training data and model weights; the full training inputs and the checkpoint-453 model export are supplied separately under `data/` and `checkpoints/`. |
|
|
| The source archive filenames and contents are preserved. Verify them with `sha256sum -c SHA256SUMS` after download. No training was started as part of this release. |
|
|
| The [clean338 checkpoint-453 model export](checkpoints/mixed_image_video_clean338_9b_32h20_gbs64_lr5e5_20260916T115800CST/checkpoint-453) is available for RL initialization under `checkpoints/mixed_image_video_clean338_9b_32h20_gbs64_lr5e5_20260916T115800CST/checkpoint-453` (11 files, 18,839,792,975 bytes). It includes four safetensors weight shards, the weight index, model configuration, tokenizer, processor configuration and chat template. Original SFT distributed training-state and optimizer checkpoint files are excluded. |
|
|
