This release: 4B, step 5949
This is the completed 5949-step SFT endpoint.
Training provenance
Both families use the registered 5949-record SFT curriculum (3898 parent groups), with 5949 maximum steps and one epoch. Each checkpoint's recorded training step, epoch, and original trainer-state hash are preserved in the separate LoRA archive manifest at fixed revision 1e5557f923746031f8187b7daf299b9bee41cb3c (repository access required). This merged package's release-manifest.json inventories inference artifacts and does not contain that per-checkpoint trainer-state record; intermediate checkpoints did not finish the whole schedule. Randomized loader order means the step number alone is not a verified count of unique examples seen at an intermediate checkpoint.
| Source | Scheduled records | Parent groups |
|---|---|---|
| Mind2Web browser Choice | 1798 | 671 |
| Mind2Web browser TYPE | 158 | 125 |
| HelpSteer3 principle | 1300 | 1300 |
| SGD | 513 | 10 |
| GoEmotions independent-attribute Score | 500 | 297 |
| BoolQ | 600 | 600 |
| MNLI | 1000 | 1000 |
| Local counterfactual | 80 | 20 |
Parent groups can overlap across browser Choice/TYPE. The 4B run initialized from a 427-record pilot adapter (weight SHA256 0047e5f1f0c98f17da93def94032041997592e1a609ddcab58de41f4d30e3a38), while the inspected 9B config records no origin adapter. Do not add pilot records to the registered schedule as if all were independent.
Decision objective: 0.5 CE(low) + 0.5 CE(high) + 0.1 KL(P_high.detach || P_low). TYPE examples use full-depth text cross-entropy. LoRA r=8, alpha=16, dropout=0; learning rate 1e-4; batch size 1; seed 20260920; training max length 6144. The recorded compiled schedule maximum is 5845 tokens. Training settings and data/schedule hashes are retained in the separate LoRA checkpoint depth configuration at fixed archive revision 1e5557f923746031f8187b7daf299b9bee41cb3c. The merged package's depth_config.json contains only portable inference and identity metadata; it is not the full training configuration.
This is SFT with a within-model distillation term; it does not prove RLCD, online reinforcement learning or teacher-model OPD occurred. The declared public sources contain multiple licensing regimes (including CC-BY, CC-BY-SA and mixed-source material); separate provenance and redistribution review remains necessary. Raw datasets are not part of this upload.