Resonance

Data for the preprint Distillation, Self-Training, and the Fragility of Learned Chemical Reasoning (Aneesh, Ibarraran, Hu, Liu, Rotskoff, 2026).

Resonance trains open-weight models to predict the product of a chemical reaction from its SMILES, reasoning step by step. A teacher (Kimi K3) writes a reasoning trace for every USPTO-50k training reaction. The student (Qwen3-8B or Qwen3.8-27B) is fine-tuned on those traces (S1), then on its own verified-correct answers (S2, rejection-sampling SFT), then on preference pairs from its own correct and incorrect answers (S3, DPO + NLL). This repository holds every dataset of that pipeline, every model's test-set outputs, and the samples behind the paper's reasoning analysis.

• Code: rotskoff-group/resonance-rxn-reasoning
• Models: anaghaa/Resonance-{8B,27B}-{S1,S2,S3,skip-S2}

Everything for this paper sits under v1/; later releases will add v2/, ... alongside it. The git tag v1 fixes this version.

Configs

Config	Rows	Contents
v1_uspto50k	40,000 / 5,000 / 5,000	USPTO-50k train / validation / test: patent id, starting material (first reactant), reagents (other reactants), product.
v1_reaction_classes	40,000 + 5,000	Rule-based reaction class per reaction, as first published (class_published) and with the benzyl-ether correction used in the paper (class).
v1_kimi_k3_traces	40,000	S1 training data: the Kimi K3 trace for every training reaction (trace, the text the students are trained on; reasoning, the teacher's hidden reasoning field). The teacher saw the product; no reaction ids were in the prompt.
v1_s2_8b	47,492	Resonance-8B S2 training set: exactly correct S1 samples (≤ 1,024 tokens) as chat-formatted text.
v1_s2_27b	98,731	Resonance-27B S2 training set (difficulty mix, including 1,108 answer-only gold fallbacks): prompt and completion decoded from the training token ids.
v1_s3_pairs_8b	28,747	Resonance-8B S3 DPO pairs: prompt, chosen (a correct S2 sample), rejected (the most Tanimoto-similar incorrect one).
v1_s3_pairs_27b	12,079	Resonance-27B S3 DPO pairs, decoded from token ids.
v1_skip_s2_pairs_8b	26,661	Skip-S2 ablation pairs, sampled from 8B S1.
v1_skip_s2_pairs_27b	13,698	Skip-S2 ablation pairs, sampled from 27B S1.
v1_test_predictions	5,000 per model	Full responses on the USPTO-50k test set for every model in the paper: both students at every stage and skip-S2, both base models, the zero-shot references (Kimi K3, Claude Opus 5, GPT-6-Astra) and the 0/4/8-shot control on base Qwen3-8B. answer is the last <answer> block, prediction its canonical SMILES, correct the paper's Top-1 exact match.
v1_reasoning_probes	27 stage × condition groups × 5,000 reactions	Resonance-8B S1–S3 samples behind the reasoning analysis: fresh resamples, the reasoning-length prompts and the feedback conditions, 8 draws per reaction (responses), with per-reaction seeds.

from datasets import load_dataset

traces = load_dataset("anaghaa/resonance", "v1_kimi_k3_traces", split="train")
preds = load_dataset("anaghaa/resonance", "v1_test_predictions", split="test")
s3 = preds.filter(lambda r: r["model"] == "Resonance-8B-S3")
print(sum(s3["correct"]) / len(s3))   # 0.428

Raw files and reproducing the paper

v1/raw/<run_dir key>/<path> holds the original files at the paths the runs wrote them, laid out for the code's resonance.paths.run_dir. The code repository fetches them:

python -m resonance.download figures   # the 63 inputs of every figure (4.3 GB)
python -m resonance.download all       # plus the USPTO split and every training set (S1, S2, DPO pairs)
python -m resonance.paper.run_all --out paper_out   # rebuilds Figs 2-7, A1, A2 on a CPU

The rebuilt figures are pixel-identical to the paper's.

Scoring

A prediction is the last <answer> block anywhere in the response. It is correct if its RDKit canonical SMILES equals that of the recorded product; invalid, missing and truncated answers count as wrong (N = 5,000). Reaction classes in the paper use the corrected labels (class).

Notes

• The reaction classes come from a rule-based classifier, not expert annotation.
• The teacher traces were written with the product visible, and some contain chemical errors; they are released as the students saw them.
• The Claude Opus 5 and GPT-6-Astra responses are zero-shot evaluation outputs, included so the paper's comparisons can be checked.
• USPTO-50k is derived from US patent data (Lowe 2012; Schneider et al. 2016).

Citation

@misc{aneesh2026resonance,
  title  = {Distillation, Self-Training, and the Fragility of Learned Chemical Reasoning},
  author = {Aneesh, Anagha and Ibarraran, Sebastian and Hu, Frank and Liu, Jason and Rotskoff, Grant},
  year   = {2026},
  note   = {Preprint}
}