Written entirely by Claude, as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post. For the readable version, see the companion.

The Dolci RL-Zero datasets (Ai2, Olmo 3)

The Olmo 3 post-training RLVR datasets -- Math, IF and Code -- used as the capability-only and style-contrast arms of the project's RL experiment.

Source

Dolci is the post-training data suite released with Olmo 3 by the Allen Institute for AI. The RL-Zero subsets are the reinforcement-learning-from- verifiable-rewards (RLVR) training data for the Olmo 3 RL-Zero 7B models. All are licensed ODC-BY and all point at the same citation:

Team Olmo et al. Olmo 3. 2025. arXiv:2512.13961.

Three are referenced in qwen35/rl_capability.py:

The Dolci collection on HuggingFace lists 32 items in total.

What this project uses them for

The zoo is 134 adapters that each teach a personality trait. qwen35/rl_capability.py trains one that teaches nothing about personality at all, to ask whether capability training displaces personality as a side effect:

Every adapter in the zoo teaches a personality trait. This one teaches nothing about personality at all: the reward is whether a maths answer is correct. If its weight delta still lands somewhere legible in the personality space the zoo defines -- loads on a principal component, moves a judged Big Five scale -- then capability training displaces personality as a side effect, and the geometry we built from traits can measure it.

The math arm uses Dolci-RL-Zero-Math with SymPy-style exact match on a scalar as the reward. Its claim to being capability-only was checked rather than assumed: the script records that all 13,314 rows were scanned, the persona-regex hit rate was 0.18%, all 24 hits were read, and there were zero true positives.

The IF arm uses Dolci-RL-Zero-IF as a deliberate style contrast: fully verifiable, and "entirely about surface form: 55.7% of its constraint instances govern diction, casing, punctuation or opening/closing wording". The pairing is the design -- one verifiable reward with no stylistic content, one verifiable reward that is almost entirely stylistic, so any weight-space movement toward the persona geometry can be attributed to style rather than to RLVR as such.

Comparability is enforced by construction. The LoRA configuration is not chosen but read off an existing zoo adapter: the same 248 module names, rank 64, alpha 128, plain LoRA, use_rslora=False, seed 0. The bilinear sketch is seeded per module name, so an adapter over the same modules projects into the same coordinates with no new plumbing. Checkpoints are saved every 25 steps and each one is sketched, so the run yields a trajectory through personality space against the reward curve rather than a single endpoint.

GRPO was chosen over rejection-sampling SFT on a preregistered pre-flight (qwen35/analysis/rl_preflight.json, 300 problems at k=8 on the base model): 101 of 300 problems have pass@8 strictly between 0 and 8, so 33.7% carry a non-zero GRPO advantage, above the 15% line registered before the measurement. The same pre-flight found boxed compliance of only 0.30 at a 3072-token cap with mean completion 2,674 tokens, so most rollouts run out of budget before answering and 33.7% is a floor.

Verification note

Dataset cards only. Both HuggingFace dataset pages were fetched 2026-09-07 for row counts, content description, source datasets, licence and citation. The Olmo 3 paper (arXiv:2512.13961) was not fetched or read -- the identifier comes from the dataset cards' citation blocks, and this page cites it on their authority. allenai/Dolci-RL-Zero-Code-7B was not checked at all -- it appears once in qwen35/rl_capability.py and its role there was not read. None of these datasets appears in qwen35/paper_notes.md; the local descriptions and the row-scan figures come from qwen35/rl_capability.py's own docstring.