tianyi0216/rlvr-selection-outputs

RLVR Selection Experimental Artifacts

This dataset repository contains the curated experimental artifacts from the RLVR selection project. It includes prompt manifests, sampled Qwen3 responses, verifier outcomes, policy-state trajectories, fitted predictions, aggregate metrics, and selected LoRA checkpoints. The source code and research handoff are available in the project repository.

The project asks whether response histories from inexpensive small models can predict success probabilities for larger RLVR policies from sparse current-policy support. The strongest completed result is a small and consistent prediction improvement on a prospectively evaluated Qwen3-14B target. A completed Qwen3-8B closed-loop study found no improvement from the tested p(1-p) selector at matched online token cost. That downstream study used the low-rank predecessor, not the final neural combination.

Release scope

This is a raw research-artifact archive. It is not one homogeneous Hugging Face table. Schemas differ by experimental stage, and the repository preserves the original directory structure under:

artifacts/Efficient_RLVR/outputs/

The release contains approximately 7.1 GB across 1,680 experiment files before release metadata. It includes 33 LoRA adapter weight files. Cloud job records, signed artifact URLs, copied execution contexts, transient logs and heartbeats, duplicate files and tar archives, and unsafe pickle-backed objects were removed before publication. See metadata/REDACTIONS.md and metadata/excluded_files.txt.

Experimental stages

Directory	Contents
e2_local_proxy and e2_local_proxy_mathcheck	Local likelihood-proxy diagnostics and mathematical checks.
e2_val_utility	Sequential validation-utility measurements.
e3_branch_utility	Common-start controlled batch-utility branches.
e3_ext	Extension from six to twenty-four controlled branches.
e3_challenge	Retrained branches evaluated on a harder validation panel.
t0_scale_transfer	Initial sealed static transfer from Qwen3-0.6B, 1.7B, and 4B to 8B.
t05_frontier_transfer	Fresh hard-math frontier transfer study.
t075_model1_bakeoff	Source-only scale-prior bake-off. The historical directory name is retained for reproducibility.
t1_source_state	Source-policy histories, predictor fits, and small-model LoRA checkpoints.
t15_neural_predictor	Expanded neural predictor bake-off.
t2_target_state_qwen3_8b	Frozen transfer to Qwen3-8B policy states.
t3_8b_selection_large_bank	Matched-cost downstream selection on a 2,048-prompt bank.
t4_neural_rescue	Calibrated policy-state prediction and prospective 14B and static 32B evaluation.
mock_e2, mock_e3, and */synth/	Software checks and synthetic diagnostics. These are not empirical results.

Common record formats

Prompt manifests usually contain a subset of the following fields:

prompt_id, dataset, category, level, question, gold,
split, row_id, source, cand_order, K_4B_screen

Early static response records commonly use:

model, revision, prompt_id, dataset, sample_id, request_seed,
response, n_tokens, finish_reason, parsed, reward, parse_ok, truncated

Later response records commonly use:

prompt_id, dataset, sample_id, reward, parsed, parse_ok,
n_tokens, truncated, seed, tag, text

Training summaries and predictor tables contain stage-specific fields for checkpoint, lineage, seed, support budget, method, NLL, Brier score, Spearman correlation, calibration error, selected heterogeneity, and mixed-group probability. Read the corresponding report in the source repository before combining tables across stages.

Example JSONL loading:

import json
from pathlib import Path

path = Path("artifacts/Efficient_RLVR/outputs/t0_scale_transfer/qwen3/responses_Qwen3-8B.jsonl")
with path.open() as handle:
    first_record = json.loads(next(handle))

Example CSV loading:

import pandas as pd

frame = pd.read_csv(
    "artifacts/Efficient_RLVR/outputs/t4_neural_rescue/t4_results.csv"
)

Model revisions

All Qwen3 experiments used dense models with thinking disabled. The recorded Hugging Face revisions are:

Model	Revision
Qwen3-0.6B	c1899de289a04d12100db370d81485cdf75e47ca
Qwen3-1.7B	70d244cc86ccca08cf5af4e1e306ecf908b1ad5e
Qwen3-4B	1cfa9a7208912126459214e8b04321603b3df60c
Qwen3-8B	b968826d9c46dd6066d109eabc6255188de91218
Qwen3-14B	40c069824f4251a91eefaf281ebe4c544efd3e18
Qwen3-32B	9216db5781bf21249d130ec9da846c4624c16137

The LoRA checkpoints use rank 32, alpha 32, and dropout 0. Adapters target the query, key, value, output, gate, up, and down projections. The language-model head is excluded. Adapter configuration files may contain revision: null; use the revision table above and the experiment path to identify the base model.

Provenance and licensing

The artifacts combine model outputs and derived records associated with several upstream resources. Their Hugging Face cards identify the following licenses as of this release:

Upstream resource	Card	License shown on card
GSM8K	openai/gsm8k	MIT
MATH	EleutherAI/hendrycks_math	MIT
DAPO-Math-17k	BytedTsinghua-SIA/DAPO-Math-17k	Apache-2.0
Qwen3	Qwen model collection	Apache-2.0

This repository therefore uses license: other. No single license is asserted over the entire mixed artifact archive. Upstream terms continue to apply to prompts, reference answers, model weights, and derived content. The project repository currently has no single top-level software or data license. See LICENSES/README.md for the release notice. This summary is provided for provenance and is not legal advice.

Many source loaders did not pin immutable dataset revisions. The released prompt manifests are the source of record for the examples used in these experiments.

Intended use

This release supports:

• Reproducing aggregate analyses from saved response and prediction records.
• Studying cross-scale success-probability prediction and sparse support calibration.
• Auditing matched-cost prompt-selection experiments.
• Inspecting selected source-model LoRA checkpoints.
• Developing new hypotheses using the existing records before paying for new generation.

The archive should not be treated as a benchmark test set, a general model-quality dataset, or evidence that current success probability is causal training utility.

Limitations

• The experiments cover dense Qwen3 models and mathematical reasoning tasks.
• Candidate prompts carry source-model observations. This is not zero-label prompt transfer.
• The T3 training bank is transductive.
• The prospective 14B panel contains three one-seed lineages with a shared initialization.
• The 32B evidence covers initialization only.
• Late 8B checkpoints have elevated truncation and are treated as stress evidence.
• Generated responses, parsed answers, and verifier labels may contain errors.
• No exhaustive manual review for incidental personal information was performed. Automated scans found no Hugging Face token, bearer credential, private-key header, or common API-key pattern in the curated release.

Integrity

metadata/SHA256SUMS contains one SHA-256 checksum for each released artifact and metadata file except the checksum files themselves. Verify a downloaded snapshot from the repository root with:

shasum -a 256 -c metadata/SHA256SUMS

metadata/release.json records the source Git revision, release rules, file count, and byte count. metadata/files.jsonl provides machine-readable path, size, checksum, stage, artifact kind, and synthetic-status fields.

Citation

A paper citation will be added when the associated manuscript is public. Until then, cite the dataset repository URL and the immutable Hugging Face revision used for your analysis.