GuoJicz518/RIDE-RL-Curriculum-v1

RIDE-RL Curriculum v1

Dataset Summary

RIDE-RL Curriculum v1 is a versioned training-target pool for reinforcement-learning post-training of RNA inverse-folding models. It contains 744 unique RNA structural targets selected from the RIDE training data and the RNA3DB train hierarchy.

The release was designed for the modular RIDE-RL training stack. A pretrained RIDE policy generates candidate RNA sequences conditioned on a target backbone, RhoFold+ predicts their structures, and structural self-consistency metrics are converted into rewards for DiffusionNFT-epsilon updates.

This is a training curriculum, not an independent test benchmark. The RIDE test split is excluded. Final model quality should be reported on a separate held-out evaluation set and should not rely on RhoFold+ as the only evaluation oracle.

Dataset Composition

Item	Count
Total targets	744
RIDE train targets	742
Novel RNA3DB targets	2
Frozen-RIDE bottom-500 intersection	113
Frozen-RIDE bottom-900 intersection	269
Passed all three oracle checks	254
Passed exactly two oracle checks	490
Short targets (L <= 50)	202
Medium targets (51 <= L <= 99)	428
Long targets (L >= 100)	114

Normalized curriculum buckets in data/targets.jsonl:

difficulty_buffer	Count
bottom_500	113
bottom_900_only	156
outside_bottom_900	473
rna3db_novel	2

The training-ready .pt manifest stores the two novel records under curriculum_source=rna3db_novel; the JSONL export also exposes them through the normalized difficulty_buffer column for convenient grouping.

Files

data/targets.jsonl
    Human- and viewer-friendly metadata for all 744 targets.

data/ride_rl_curriculum_production.pt
    Training-ready PyTorch manifest with target metadata, reference coordinates,
    masks, native-refold calibration, and frozen-RIDE screening annotations.

metadata/manifest_summary.json
    Release configuration, aggregate counts, and per-target metadata.

metadata/verification.json
    Machine-readable publication audit. All release checks are true.

metadata/checksums.sha256
    SHA256 checksums for all published data files.

SELECTION_PROTOCOL.md
    Detailed English selection pipeline and acceptance rules.

Selection Overview

1. RIDE training targets

1. Start from 4,025 records in the RIDE train split.
2. Require a runnable sequence length, canonical nucleotides, finite coordinates, and complete required backbone atoms.
3. Keep 1,700 strict-complete targets.
4. Generate two sequences per target with the frozen pretrained RIDE policy.
5. Rank targets by mean sequence recovery within length strata and annotate bottom-500 and bottom-900 difficulty buffers.
6. Do not accept a difficult target solely because RIDE performs poorly; it must also pass the RhoFold+ native-refold reliability gate.

2. Novel RNA3DB targets

1. Use only the RNA3DB 2026-01-05 single-chain train hierarchy.
2. Require length 32..280, canonical or safely normalized modified bases, acceptable resolution, complete required atoms, finite coordinates, and bounded terminal trimming.
3. Remove duplicate sequences and duplicate structure hashes.
4. Exclude near homologs of all 4,223 RIDE processed sequences when LCS / longer_length >= 0.80.
5. Apply the same RhoFold+ native-refold reliability gate used for RIDE targets.

The RNA3DB archive contained 13,451 train chains and 1,990 test chains. The quality and novelty filters produced 37 candidates; 2 passed the strict oracle gate and entered this release. Both were released in or after 2024.

3. RhoFold+ native-refold gate

The combined candidate pool contained 1,737 targets: 1,700 RIDE train targets and 37 novel RNA3DB candidates.

For each target, its native sequence was folded with RhoFold+ and compared with its target structure using the same C4-prime, same-index Kabsch metric implementation used by the RL reward. A target is accepted when at least two of three checks pass:

Check	Threshold
TM-score	>= 0.45
GDT-TS	>= 0.50
RMSD	<= 2.0 Angstrom

Results:

• 1,733 targets produced valid native-refold metrics.
• 4 targets hit the 300-second worker timeout and were excluded.
• 744 passed at least two checks and were released.
• 993 failed the strict gate and were excluded from the production pool.

The gate measures whether the reward oracle can reliably refold the native sequence toward the target. It is not the trained policy's evaluation score.

Validation Rules

The publication audit verifies that:

• the pool contains at least 500 targets;
• every record loads through the typed RIDE-RL target loader;
• target IDs, sequences, and structure hashes are unique;
• all 744 calibration records have status ok;
• all 744 targets have oracle reliability tier reliable;
• the pool contains underperforming RIDE train targets;
• the pool contains strictly reliable new RNA3DB targets;
• selected RNA3DB targets have no RIDE near-homology hit at the 0.80 threshold;
• no RIDE test target is included.

Loading the Dataset

Metadata-only access:

from datasets import load_dataset

dataset = load_dataset("GuoJicz518/RIDE-RL-Curriculum-v1", "metadata")
print(dataset["train"][0])

Training-ready manifest:

import torch
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="GuoJicz518/RIDE-RL-Curriculum-v1",
    repo_type="dataset",
    filename="data/ride_rl_curriculum_production.pt",
)
pool = torch.load(path, map_location="cpu", weights_only=False)
print(pool["schema_version"])
print(len(pool["targets"]))  # 744

Only load PyTorch pickle-based manifests from a trusted repository and pinned revision. The JSONL file is sufficient for metadata inspection.

With the RIDE-RL repository:

export RIDE_RL_POOL=/absolute/path/to/ride_rl_curriculum_production.pt
python train_rl.py \
  --config configs/rl_curriculum_v1.yaml \
  --mode train \
  --output-dir outputs/rl_curriculum_v1/train

The checked-in configuration intentionally uses a one-step/eight-oracle-call safety budget. Increase the training budget only after the environment smoke test passes.

Intended Uses

• RL post-training of RIDE or compatible structure-conditioned RNA sequence models.
• Curriculum sampling experiments using source, length, or difficulty annotations.
• Reproducible oracle-reliability and data-selection studies.
• Development of reward, rollout, and advantage modules on a frozen target set.

Limitations and Biases

• Selection is explicitly biased toward targets that RhoFold+ can refold reliably.
• Only 2 of 744 targets are novel RNA3DB additions under the strict first-release gate.
• Frozen-RIDE difficulty is based on two generated sequences per target and is a screening signal, not a precise estimate of target difficulty.
• Rfam and RNA type labels inherit the quality and coverage of the source records.
• The pool is not suitable as a held-out benchmark for models trained on it.
• DRFold was not used in selection or validation.

Sources and Licensing

The release derives target records from the RIDE processed training data and the RNA3DB 2026-01-05 train hierarchy. It does not relicense upstream structural records. Users are responsible for complying with the applicable RIDE, RNA3DB, wwPDB/PDB, and source-entry terms. See:

• RIDE upstream repository
• RIDE-RL code and reproducibility files
• RNA3DB 2026-01-05 release
• RhoFold+ repository

Citation

If this curriculum is useful, cite the upstream RIDE work and the source datasets. The RIDE citation is:

@inproceedings{hu2026rider,
  title={{RIDER}: 3D {RNA} Inverse Design with Reinforcement Learning--Guided Diffusion},
  author={Hu, Tianmeng and Cui, Yongzheng and Luo, Biao and Li, Ke},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026}
}

Release

• Schema: ride_rl_target_pool.v1
• Release: v1
• Manifest creation time: 2026-08-24T03:58:16.922621+00:00
• Production manifest SHA256: 6560437498c225c0ee14c8ca06dd0796be75a8877baee198b5394895ae503ac1