Selected HASE checkpoint recovery
This repository preserves ten selected model assets from the original HASE experiments and the October 1 controlled rebuttal comparison. It is a recovery archive, not a new benchmark evaluation. Checkpoint selection follows recorded paper results, actual evaluation-loading logs and meaningful control endpoints, not every training step. No new training, fusion or inference was performed.
| Asset | Purpose |
|---|---|
s2d_submission_p2_fused |
Actual inference export used for the paper's 86.98 ± 0.38 S2D result |
alpha_submission_phase5_step16 |
Original alpha Phase 5, identified by the training save log |
k20_full_phase3_step9 |
October 1 Full HASE Phase 3 endpoint |
k20_weights_phase3_step9 |
Matched fixed-harness Phase 3 endpoint |
circle_ratchet_success_step94 |
Successful evaluator-repair trajectory's paper-level node |
k20_full_phase2_step6 |
October 1 Full HASE Phase 2 control |
k20_weights_phase2_step6 |
Matched fixed-harness Phase 2 control |
s2d_submission_p2_training_source |
Original training source of the paper's p2 inference export |
circle_n32_transfer_step130 |
Meaningful recorded N=32 transfer endpoint |
s2d_success_alternative_p3_fused |
Additional successful S2D inference export |
Each asset is under checkpoints/<asset_id>/. _HASE_ASSET.json gives provenance,
original relative path and the file manifest. CHECKPOINT_INDEX.json maps files
to original restore paths and SHA-256 hashes. SHA256SUMS verifies a complete
download; individual assets can be downloaded instead of the whole archive.
The total selected source size is approximately 282.62 GiB. FSDP training sources retain model, optimizer and extra-state shards plus configuration/tokenizer files; inference exports retain their original safetensors and tokenizer artifacts. Restore training shards using the declared world size and matching VeRL version. The root of this multi-asset repository is not a single Transformers model.
The original alpha step16 directory contained obsolete five-rank shards alongside
the later four-rank save. Only the complete set matching fsdp_config and the
recorded final save was archived; no original files were deleted.
Source recovery and VeRL patch instructions are in the GitHub rescue branch. Original research records and financial inputs are preserved separately in a private archive, not published here. Hualin-hosted experiments are outside this machine-reclamation backup scope. Applicable upstream model/component licenses continue to apply; this archive does not assign one blanket license to all components.