ModeBench and ReplayMaxRL · Research artifacts
Terminal fine-tuned models for studying correctness and verified output support. Browse by study, base model, domain, method and seed; restore each export from its recorded immutable revision.
Paper artifact collection · 532 / 707 selected exports verified. Transfer is in progress.
The fixed archival selection contains admitted completed models from the source studies listed below. Counts describe archive availability, not live training progress or a replacement for the papers’ frozen analysis sets.
The broader paper collection is being expanded to additional baselines, ablations and supporting data. This page reports only the listed verified exports; completion of this selection does not imply that every paper artifact has been archived.
Models · Benchmarks · Data · Results · Reproducibility
| Study | Question and model set | Verified / selected |
|---|---|---|
| E78 | Verified replay · Qwen2.5-0.5B | 50 / 50 |
| E79 | Verified replay · Falcon3-1B | 49 / 49 |
| E80-R1 | Verified replay · Qwen2.5-3B | 50 / 50 |
| E118 | MaxRL × verified replay · Level 1 | 138 / 139 |
| E119 | Four-method factorial · Level 2 | 88 / 89 |
| E120-R1 | Uniform versus fresh-frequency replay | 43 / 44 |
| E95 | Plain GRPO · initial cross-scale cohort | 0 / 0 |
| E114 | Plain GRPO · Qwen2.5-3B seed extension | 20 / 20 |
| E97 | UCPO · Qwen2.5-0.5B initial domains | 0 / 15 |
| E99 | UCPO · Falcon3-1B | 0 / 25 |
| E115 | UCPO · domain and scale extension | 0 / 10 |
| E98-R1 | Sparse RLEP-Dr · Qwen2.5-0.5B | 0 / 15 |
| E100 | Sparse RLEP-Dr · Falcon3-1B | 22 / 22 |
| E116 | Sparse RLEP-Dr · domain and scale extension | 0 / 10 |
| E81 | Fixed semantic + replay · Qwen2.5-0.5B | 0 / 20 |
| E82 | Fixed semantic + replay · Falcon3-1B | 25 / 25 |
| E83 | Fixed semantic without replay · Qwen2.5-0.5B | 0 / 20 |
| E85 | PantryPlan semantic-interface repairs | 0 / 10 |
| E86 | Fixed semantic without replay · Falcon3-1B | 13 / 25 |
| E87 | Historical fixed-semantic probe · Qwen2.5-3B | 5 / 5 |
| E109 | Repaired Python replay comparators | 5 / 10 |
| E112-R1 | Verified-support semantic + proposal bundle | 24 / 49 |
| E121 | Fixed-bank exemplar survival | 0 / 5 |
Artifact availability
| Scientific records selected | Deployable model exports | Verified model exports | Model weights unavailable |
|---|---|---|---|
| 762 | 707 | 532 | 55 |
Model weights are unavailable for 55 retained scientific records: E95 (55). Their metadata and admitted results remain part of the research record; they are excluded from the downloadable-model catalog.
Study and comparator map · Restore a model · Licenses · Full catalog
Find the models behind a result
- Retention across model scales: E78, E79, E80-R1.
- MaxRL × replay, Level 1: E118 and its original Dr.GRPO comparators.
- Matched Level-2 factorial: E119.
- Uniform versus fresh-frequency weighting: E120-R1 and its 45 original uniform comparators.
The older comparator exports keep their E78/E79/E80-R1 names and appear once in the catalog. Training restarts do not create extra scientific cells. Model availability alone does not establish an admissible paired analysis; use the registered seed intersections and source checks in the papers.
Restore one model
Open a study table and select a verified model. Each link fixes the full upload commit, while catalog.json provides its complete commit_sha and repo_prefix. Download only that folder with snapshot_download(repo_id=..., revision=commit_sha, allow_patterns=[repo_prefix + "/*"]) and verify the original file sizes and SHA-256 values in its ARCHIVE_MANIFEST.json.
Exports include weights, configuration and tokenizer assets. They restore inference models; optimizer state and online replay banks are not included. Preserve the original domain prompt format, generation settings and verifier when reproducing a result.
Base models and licenses
| Base model | License and attribution |
|---|---|
| Qwen2.5-0.5B-Instruct | Apache License 2.0; original Alibaba attribution and modification notices retained. |
| Falcon3-1B-Instruct | TII Falcon-LLM License 2.0 and incorporated acceptable-use policy. |
| Qwen2.5-3B-Instruct | Qwen Research License Agreement, including its research/evaluation scope. |
Built with Qwen. The Falcon3-1B experimental models in this archive are built using artificial intelligence technology from the Technology Innovation Institute. Each model folder carries its applicable LICENSE, MODIFICATIONS.md, and any required Notice or ACCEPTABLE_USE_POLICY.html.