Perfect7613/jump-gemma-adapter

Track H Gemma adapter artifacts

This sanitized repository contains adapter-only weights from two bounded Track H runs. The base checkpoint is google/gemma-4-12B-it@707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7; base weights are not included. Transformers is pinned to official source commit 918dbf131d0df5b46e3f6e1d96174d62aa4d16d6.

The real pilot trained 1,298,432 LoRA parameters per condition on 200 deterministic episodes and evaluated 24 held-out episodes. Primary answer-token accuracy was 0.530; none of four generated primary answers parsed as complete JSON. The secondary runs are text ablations, not the PRD-locked no-hidden-world or permuted episode–latent controls. They provide no G3 or mechanistic evidence.

Conditions

• primary: true episode partition serialized in the prompt.
• types-unavailable-text-ablation: literal type-unavailable prompt. This is not a latent learned from no-hidden-structure worlds.
• wrong-partition-text-ablation: within-episode shuffled partition text. This is not permuted episode–latent training.

All conditions used identical examples, order, 25 steps, optimizer, and LoRA topology. Base trainable parameters were zero. See the manifests for the frozen definitions and recovery lineage.

Intended use and limitations

These artifacts support reproduction and inspection of this exact synthetic benchmark run. They are not general-purpose adapters. Exact structured generation was 0/4 in every condition, the pilot has one seed, and the ablations do not satisfy the planned training-control definitions. The results do not establish a hidden representation, causal effect, or mechanism.

The pinned upstream model metadata reports Apache-2.0, while its model card also links the separate Gemma 4 License. Use of these adapters requires reviewing and complying with the upstream model terms as well as the source project license posture. Public availability of adapter-only artifacts does not itself establish rights to the base model; users must comply with the linked upstream terms. Only JUMP-produced adapter weights and metadata are included. No optimizer state, cache, credential, base weight, or third-party tokenizer file is present.

Reproduction references

Use the exact base and Transformers revisions above, then load a condition from pilot/adapters/<condition>/<condition> with PEFT. pilot/training-manifest.json, pilot/dataset.json, pilot/tokenization-audit.json, pilot/result.json, and checksum files bind the run. The tokenizer is referenced from the pinned base rather than copied.

Exploratory endpoint limitations

This adapter was selected for JSON parse validity, not answer correctness. Its one-time final test produced valid JSON for 12/12 examples and exact partition-plus-law answers for 3/12. Validation and final test each contain 12 examples, so their rates have high uncertainty.

The synthetic generator used only four law combinations and confounds the same-type sign, different-type sign, and exponent. Law accuracy from this artifact is not benchmark, G1, or G3 evidence. A confirmatory generator must randomize those components independently.

Step 50 was selected on validation data by maximizing exact accuracy, then maximizing parse rate, then minimizing teacher-forced loss. The final test was evaluated once after restoring step 50. The endpoint applies the same schema-derived grammar used in validation and final testing. Grammar validity does not establish correctness.

Privacy and publication status

This repository was recreated from the independently verified sanitized current tree. No commits, discussions, live endpoint descriptors, or other history from the private archive were migrated. The fresh history and current files were scanned before public visibility.