deeprs-adapters

deeprs — LoRA adapters

Every adapter trained for Amplification without Acquisition: what reinforcement learning buys when training a language model to analyse scientific data, and which capability the training spends.

All adapters are LoRA on Qwen/Qwen3-8B unless the name says 14b.

The headline model

tenrl-step150-lora scores 0.994 across all ten public GeneBench-Pro problems — 8 of 10 at exactly 1.000 over 256 rollouts each — where the same base model untrained scores 0.014, and the best published frontier system reaches 36% on that split.

Why all 120 are here

The paper's load-bearing claims are comparisons. The transfer result is that every leave-one-out model scores below the untrained base; the forgetting result compares four corpora; the stability result is that identical corpora give 0.029 or 0.341 depending only on seed. Verifying any of those needs the specific adapter for that arm, so publishing only the winner would leave every negative result unverifiable.

Reproducing

Code, paper draft, and a results index that recomputes every published number: https://github.com/fatebreaker/deeprs — start at CONTINUITY.md.

The problem instances these were trained on are in a separate, access-gated dataset repository; note that re-running the generators does not reproduce them, as the generators were edited after the corpora were built.