Gemma4 equiangular-line repair adapters

Fresh strict-jammed SFT and continued RL adapters for exact fixed-angle equiangular-line construction/repair. These are structured JSON action policies, not general-purpose chat models. The release follows the Gemma4 Salem-Spencer layout: sft/, followed by rl/checkpoint-N/ adapters and tokenizer files.

Reproduce and continue evaluation

Clone the jammed-eql branches of extremal and llmcosolver. Follow the complete evaluation handoff. It includes pinned downloads, exact CPU verification, reference reconstruction, software requirements and a portable torchrun evaluator. Code revision: c94246ab06458724194820095f3ca92b3f6faabc (the later release-pin commit only adds immutable Hub references).

Data: seanmamasde/extremal, revision 938e2e00d5b1590e81ddc3d4e0abcf65445887b2. File SHA256s and inference settings are in release.json.

Required model dependency chain

Raw base -> merge SFT -> load RL adapter.

  • Raw base: google/gemma-4-26B-A4B, revision 24548b62aa021d562695c04aaf7758a1ea47990b (pretrained, non-IT).
  • Load raw weights in BF16, load the released SFT adapter in FP32, and use PEFT merge_and_unload(safe_merge=True) on CPU, as the original training did.
  • Apply the selected FP32 RL adapter to that frozen SFT-merged reference.
  • Both SFT and RL adapters here are FP32; the SSSET SFT adapter used a different dtype. Do not copy its BF16-adapter merge assumption into this experiment.
  • RL adapter configs intentionally do not select a raw base automatically. Applying them directly to the raw Google model skips the SFT update.

The upstream raw model has its own Apache-2.0 terms. Its weights are not duplicated here. Re-merging and inference can differ in low-order bits across hardware and numerical backends; fixed seeds do not guarantee bitwise-identical trajectories.

Checkpoints

Directory Completed work Evaluation status
sft/ 4558 updates, one full-batch pass Frozen reference for every RL adapter
rl/checkpoint-580/ RL round580 Frozen control under the latest OOD protocol
rl/checkpoint-600/ RL round600 Full120-case suite
rl/checkpoint-625/ RL round625 Latest72-case OOD/frontier suite
rl/checkpoint-632/ RL round632, Adam1716 Latest saved policy; not yet evaluated

training_state/checkpoint-632.pt additionally preserves the final full adapter and Adam state. It is optional for evaluation. Historical paths and reference identities inside that state are provenance; a training continuation must explicitly validate/rebind identities and retain optimizer moments.

Data and search contract

SFT contains145882 training actions and15704 validation actions from exact strict-jammed repair trajectories. The completed SFT pass used145856 examples, with26 outside the last full global batch. Training permits d2-17, with actual strict-jammed roots in d4/5/14/15/16/17. The repeated-repair bank has5906 training trajectories. Held-out geometry checks separate validation, and d18-20 OOD never enters optimizer updates. Synthetic/expert actions are excluded from PPO.

The latest fixed OOD bank has26 geometries and two seeds each:32 d18,10 d19 and 10 d20 attempts. Additional solver frontier starts are reported separately and rotate through four slots. OOD uses2048 expansion calls,8 proposals/call, resamples4,32768 nodes, a192 charged-turn path budget and96-line capacity. Train/validation use1024 calls. Context is8192 tokens; generations use32 tokens with temperature1, top_p1 and top_k disabled. There is no evaluation walltime cap.

The primary OOD metric requires exceeding the published57/72/90 lower bounds, so the breakthrough thresholds are58/73/91. Local improvement beyond a root's starting size is not a published-bound advance and is not a search stop condition.

Observed results

The latest evaluated policy,625, has8/52 local OOD repairs and zero published-bound improvements. Against600 on the same fixed bank it has one cardinality win, 50 ties and one loss. The frozen580 control under the same OOD budgets also has zero breakthroughs and a larger total local line gain than625. Checkpoint632 has no measured search result yet.

eval_results/ contains compact saved results for these comparisons, including exact returned witnesses. Full original prompts/trees remain in the source research archive; they are not included in the compact published result files. The portable evaluator records complete traces for new runs.

Reference software: Python3.12, Torch2.11.0+cu129, Transformers5.14.1, PEFT0.20.0, vLLM0.26.0+cu129, pynauty2.8.8.1 and Gymnasium1.2.3. The handoff provides exact installation pins and the official CUDA12.9 vLLM wheel URL.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for seanmamasde/gemma4_equiangular

Adapter
(3)
this model