Despaireyes613/learning-without-looking

Same Reward, Different Skills: data and evaluation outputs

This repository holds the data behind the paper Same Reward, Different Skills. The paper studies what reinforcement learning with verifiable rewards teaches a vision-language model about looking, and shows that training problems whose answers require the image teach visual discovery.

Everything here is consumed by the code repository, https://github.com/yhc98002-bit/learning-without-looking. Its scripts/fetch_data.py downloads the parts you need. The default part (about 22 MB) holds the per-item outputs together with the training streams and audit rows the analysis reads, and scripts/reproduce.py rebuilds the paper's tables and figures from it on a CPU. A handful of printed values come from sources outside the per-item outputs, such as the trainer's reward log; the rebuild marks them as not rebuildable.

Contents

path	size	what it is
predictions/	11 MB	per-item evaluation outputs of every run the paper reports, one folder per experiment family
data/scenes/	252 MB	the constructed coordinate scenes: training, development and confirmatory sets, each scene rendered as a counterfactual twin pair with four cue levels
data/grounding/	173 MB	the held-out grounding instruments: a 1,200-pair counterfactual suite and its independently regenerated twin
data/training/	104 MB	training rows for every run: the constructed corpus, the three dose mixtures, the filtered Geometry3K rows and images, the filtered ViRL39K rows
data/audit/	189 MB	inputs to the visual-necessity audit: the constructed rows and a 4,096-item ViRL39K sample with its 4,297 images
data/captions/	46 MB	question-blind caption stores: the 3B captions used by the caption-condition arms, and the 72B captions of the grounding instruments used for the caption-stress check
data/benchmarks/	8 KB	the benchmarks of the necessity audit, with the evaluation harness revision; the benchmark data itself is not redistributed
checkpoints/	16 GB	the 7B model trained on the constructed corpus with the standard reward (run 1, step 100)

Archives (*.tar.gz) unpack in place, so data/scenes/training.tar.gz becomes data/scenes/training/. python scripts/fetch_data.py --part instruments --part corpora --extract downloads and unpacks them.

Per-item outputs

Each folder under predictions/ holds up to three gzipped JSON-lines files.

• pairs.jsonl.gz has one row per counterfactual pair evaluation. The identity fields are pair_id, eval_set, task, layer (l1, l2, l3 or probe), density and role. The run fields are run, measurement, model, train_condition, seed, step and test_condition. Each member carries gold_a/gold_b, the recorded answer span prediction_a/prediction_b, and the extracted answer answer_a/answer_b. Correctness is in correct_a, correct_b, pair_correct, pair_correct_strict and collapsed. These fields come from rescoring the recorded spans with the repository's scorer. The *_as_logged fields keep the flags recorded at evaluation time, which for the earliest runs used an earlier matcher.
• items.jsonl.gz has one row per single-image item. correct is the canonical answer matcher and correct_reward_matcher is the training reward's matcher. Benchmark and audit-sample items also carry answer_format, n_choices, and for multiple choice the option_labels shown and the gold_labels, which the chance-corrected retention needs.
• audits.jsonl.gz has one row per item of a 16-sample audit: samples, then correct_samples and p_sample under both matchers, plus the greedy result. The resolvability audit of the training corpora uses the training reward's matcher (*_reward_matcher); the benchmark audit uses correct.

measurement says which evaluation of a cell a row belongs to, and each table reads a named one:

• primary is the evaluation reported for a run.
• scene-baseline, instrument-release, twin-release, remeasure, scale-study and long-horizon-stem label the baseline and instrument evaluations that particular tables use.
• caption-stress-3b-captions and caption-stress-72b-captions are the caption-only checks.
• scene-baseline-first-render and caption-stress-72b-captions-first-render were run on the first rendering of the development scenes.
• early-read is an earlier evaluation of the step-20 checkpoint of the standard 7B run; the tables use the later one.
• repeat is a restarted evaluation whose outputs match the primary one.

predictions/manifest.json lists the record counts, per-cell counts and SHA-256 digests.

Licences and attribution

• The constructed scenes, the grounding instruments and the per-item outputs are released under Apache-2.0.
• Geometry3K (MIT) supplies the Geometry3K rows and images.
• ViRL39K (MIT) supplies the ViRL39K rows, including items whose source is MMK12 (Apache-2.0). The full ViRL39K image pool is not redistributed: the training rows record each image's SHA-256, and the code repository explains how to place the upstream images. Only the 4,096-image audit sample (4,096 items, 4,297 images) is included.
• Built with Qwen. The 72B captions were generated by Qwen2.5-VL-72B-Instruct and carry the Qwen License. The 3B captions were generated by Qwen2.5-VL-3B-Instruct and carry the Qwen Research License, which restricts them to non-commercial use.
• The checkpoint is a fine-tune of Qwen2.5-VL-7B-Instruct (Apache-2.0).
• Benchmark data used in the audit (MMVP, BLINK, MathVerse, MMMU, HallusionBench, MathVista, MMStar) is not included; it is read through VLMEvalKit under each benchmark's own terms.
• predictions/necessity_audit_cross_family holds the answers of Gemma-3-12B-it and InternVL3-9B on the ViRL39K audit sample and on the grounding suite and twin, which are subject to those models' terms.