jgaeb/rubric-rewards

Replication Data for "From Constitutions to Control: Interpretable Rewards for Aligning Language Models"

This repository contains all replication data for "From Constitutions to Control: Interpretable Rewards for Aligning Language Models"; replication code is available at github.com/jgaeb/rubric-rewards, and the fine-tuned adapters are available from the model repository jgaeb/rubric-rewards-adapters.

Data are available at three levels of processing:

• raw/: Raw data stored as JSONL. These include every prompt, candidate response, rubric grade and the grader's reasoning, elicited ranking, fitted reward, fine-tuning set, policy generation, and judgment.
• tables/: the same records as parquet tables, one row per observation, built from raw/ by build_tables.py. Every figure in the paper is drawn from these tables.
• figures/: the paper's figures, drawn from tables/ by plots.R, each with the data it plots as one or more CSV files.

MANIFEST_raw.json lists the SHA-256 digest of every file in raw/.

Organization

Data tables share the following columns:

• model: the base model, one of gemma (google/gemma-3-12b-it), qwen (Qwen/Qwen3-14B), and gptoss (openai/gpt-oss-20b).
• reward: free (the unconstrained Plackett-Luce fit, used in the main text), constrained (the sign-constrained fit), discrimination (the policies of the label-bias experiment), or base_reasoning_off (GPT-OSS's base model answering in its final channel, the reference for GPT-OSS).
• arm: one fine-tuned policy. help0.5_lean-1 has β_help = 0.5 and β_lean = −1; helpinf_lean0 and help0_lean+inf are lexicographic; syc0 through sycinf are the sycophancy penalties; skywork8b_q0.4 was trained on the picks of Skywork 8B fine-tuned at injected bias q = 0.4; and base is the untrained model.
• lever: feedback (sycophancy), harm, political, jobs (discrimination), movies, mmlu, and ifeval.
• split: fit (fits the reward), train (fine-tuning), eval_in (in-distribution evaluation), and eval_out (out-of-distribution evaluation).
• rm and scorer: the reward model, one of rubric, rubric_constrained, skywork8b, skywork3b, skywork1b (the three Skywork-Reward-V2 models), qrm (nicolinho/QRM-Gemma-2-27B), and deberta (OpenAssistant/reward-model-deberta-v3-large-v2).

Tables

table	one row per	additional columns
rubric_items	rubric item (329)	family, subsection, direction, active (89 items), graded_by
rubric_structure	rubric subsection (50)	n_items, n_kept
prompts_<lever>	prompt	prompt_key, split (none for mmlu and ifeval), prompt, and the lever's metadata
responses	candidate response	model, key, prompt_key, lever, split, response, reasoning, and the member_* columns describing how the response was generated
grades	fully graded response	one column per active item holding its grade
rankings	(pool, rank)	pool_key, rank, key, in_reward_fit
weights	(fit, rubric item)	reward, q (empty except for the discrimination refits), weight, se
sft_targets	fine-tuning example	key (the selected response, or tulu:i for a Tülu-3 example), source
policy_judge_obs	(arm, prompt, metric)	metric (harm 1–5, help 0–3, political_slant −3 to 3), value, lean_label
policy_syc_obs	(arm, writing sample)	rating_like, rating_dislike, syc_framing_gap
policy_capability_obs	(arm, question, metric)	metric (mmlu_acc, ifeval_strict, ifeval_loose), value (0 or 1)
policy_generations	(arm, evaluation slice, prompt)	slice, response, reasoning, finish_reason
fleet_judge_obs, fleet_syc_obs	(comparison model, prompt)	the 19 comparison models, scored like the policies
disc_jobs_obs	(arm, hiring case, name)	pick, aligned (1 if the pick follows the gender stereotype)
disc_movies_obs	(arm, movie pair)	choice_man, choice_woman, congruence
rm_scores	(response, reward model)	scorer, score, in_reward_fit
rm_correlations	(model, lever, pair of reward models)	spearman, n

Tables of trained policies also carry the columns that identify the arm: model, reward, arm_type (base, grid, syc, or dose), beta_help, beta_lean, beta_syc, rm, q, and arm.

Raw records

• rubric/: the 329 rubric items, the grading prompts with their worked examples, the penalty groups, and the full text of the constitution with its publication date and digest.
• prompts/: every prompt, by lever.
• responses/, grades/, rankings/: each model's candidate responses, their grades with the grader's reasoning, and the elicited rankings.
• rewards/: the fitted rubric weights, including the discrimination refits, and <model>_fit_cells.jsonl, the grades the rewards were fit on where they differ from grades/ (78 Gemma probe grades redone after the fit, and the grades of 7 GPT-OSS responses that are not fully graded), which refit.py --exact uses.
• sft/: every fine-tuning set, including the 1,000 examples the GPT-OSS softened-response generator was trained on.
• policy/: the fine-tuned policies' generations, judgments, and MMLU and IFEval results.
• fleet/: the comparison models' generations and judgments.
• discrimination/: the preferences at each injected bias, the reward-model training pairs, every dose reward model's scores, the fine-tuning sets, and the dose policies' generations.
• rm_scores/: the five neural reward models' scores on every candidate response.
• prompts_used/: every prompt template used to generate, grade, rank, and judge.

Sources and licenses

The data build on PKU-SafeRLHF (Ji et al., 2025), the feedback task of Sharma et al. (2024), MovieLens 1M, the BLS Current Population Survey (Table 11, 2025), the Tülu-3 SFT mixture, MMLU, IFEval, and Claude's constitution (Anthropic, published January 21, 2026, reproduced in raw/rubric/ because the rubric and every elicitation prompt depend on its exact text). Each keeps its own license. Our annotations, prompts, and generations are released under CC-BY-4.0.