| # posttrain-ckpt-2000 — bounded post-train of the released T-Rex midtrain checkpoint |
|
|
| Weights of the 2000-step post-training run `trex_posttrain_public_0728_1300` |
| from https://github.com/physercoe/trex-repro (see `docs/EVAL.md`). |
|
|
| - **Base**: officially released T-Rex *midtrain* checkpoint (upstream: |
| https://github.com/ZhuoyangLiu2005/T-Rex; arXiv:2606.17055). |
| - **Data**: 144 episodes (shard 0) of the public `zekaiwang/trex_dataset`, |
| adapted on-the-fly to the eef-62 schema (`scripts/trex_public_dataset.py`). |
| - **Config**: MAX_STEPS=2000, bsz 4, lr 1e-4, bf16; train loss 1.48 → |
| ~0.07–0.23; val @1999 act=0.0588, tac=0.1152. |
| - **Files**: `model.pt` (single-file torch checkpoint, ~8.5 GB — weights only, |
| no optimizer state), `config.json`, `training_args.json`, `stats_data.json`, |
| `processor/` (tokenizer/preprocessor). Load with the repo's own |
| `model_load` (`repo/scripts/test.py`) — the format round-trips through its |
| auto-detect. |
| - **Finding** (candid): open-loop, this post-train did not make the tactile |
| expert input-responsive (probes ~5e-4 before and after) — see the repo |
| README and `docs/EVAL.md`, Finding P3. |
|
|
| License: MIT. Derived from the T-Rex authors' released checkpoint — cite |
| arXiv:2606.17055. |
|
|
