episode_id int64 0 19 | topic stringlengths 9 30 | keyword stringlengths 3 10 | W int64 8 8 | documents listlengths 8 8 | qas listlengths 236 262 |
|---|---|---|---|---|---|
0 | new nebula | Prysm-4 | 8 | [{"body":"The discovery of Prysm-4, a young, expanding planetary nebula located in the Cygnus-X comp(...TRUNCATED) | [{"agreement":"exact","answer":"Helios-9 Interferometric Network","answer_source":"original","answer(...TRUNCATED) |
1 | new metamaterial | Chi-9 | 8 | [{"body":"The design and fabrication of acoustic metamaterials have increasingly shifted from simple(...TRUNCATED) | [{"agreement":"exact","answer":"20 mm","answer_source":"original","answer_type":"short","arity":"sin(...TRUNCATED) |
2 | new virus | Jaw | 8 | [{"body":"The emergence of the novel viral agent designated \"Jaw\" in late 2024 necessitates an imm(...TRUNCATED) | [{"agreement":"exact","answer":"0.1","answer_source":"original","answer_type":"short","arity":"singl(...TRUNCATED) |
3 | new ceramic compound | Graphenex | 8 | [{"body":"The development of advanced ceramic compounds capable of withstanding extreme thermomechan(...TRUNCATED) | [{"agreement":"exact","answer":"50 µm","answer_source":"original","answer_type":"short","arity":"si(...TRUNCATED) |
4 | new plant | Montia | 8 | [{"body":"The discovery of *Montia*, a novel photosynthetic organism isolated from the deep crevices(...TRUNCATED) | [{"agreement":"exact","answer":"CC-9000","answer_source":"original","answer_type":"short","arity":"s(...TRUNCATED) |
5 | new element | Yolk | 8 | [{"body":"The discovery of Yolk-238 has necessitated a rigorous immediate examination of its fundame(...TRUNCATED) | [{"agreement":"exact","answer":"50 microamperes","answer_source":"original","answer_type":"short","a(...TRUNCATED) |
6 | new ceramic compound | Westex | 8 | [{"body":"The development of Westex represents a pivotal advancement in high-temperature ceramic com(...TRUNCATED) | [{"agreement":"exact","answer":"1.5:1","answer_source":"original","answer_type":"short","arity":"sin(...TRUNCATED) |
7 | new nebula | Gfr | 8 | [{"body":"The discovery of the Gfr nebula represents a paradigm shift in understanding high-energy s(...TRUNCATED) | [{"agreement":"f1","answer":"R=100,000","answer_source":"original","answer_type":"short","arity":"si(...TRUNCATED) |
8 | new insect | Jylix | 8 | [{"body":"The discovery of *Jylix vespiformis*, a previously undocumented insect belonging to the or(...TRUNCATED) | [{"agreement":"exact","answer":"Hemiptera","answer_source":"original","answer_type":"short","arity":(...TRUNCATED) |
9 | new alien organism | Seal | 8 | [{"body":"**Abstract**\nThis study establishes the foundational morphological and physiological para(...TRUNCATED) | [{"agreement":"exact","answer":"Zeiss Sigma 500","answer_source":"original","answer_type":"short","a(...TRUNCATED) |
synthetic_science_v3 — sample
20 episodes (160 documents, 5,022 QAs) from the v3 corpus, plus the flattened single-document rows the meta-TTT trainer actually consumes.
Three configs
episodes — the corpus as generated. One row = one episode: episode_id, topic, keyword, W, documents[], qas[].
documents[i]:{idx, title, body, summary, n_tokens}— ~1,789 tokens eachqas[i]:{question, answer, style, arity, source_doc_ids, doc_gap, answer_type, choices, cot, cot_answer, original_answer, answer_source, agreement}
single_doc_train_rows — one row per (document, QA). This is the training
format: passage_id, passage, question, answer, short_answer, qtype.
eval_heldout_rows — same schema, held-out episodes, no summary QA.
Verification trail
Every QA carries how its gold was decided. A CoT was derived blind (without showing the model the generated answer), then compared:
original_answer— from the QA-generation stagecot_answer— from the blind derivationanswer_source—"original"or"cot"agreement—exact/f1/llm_same/llm_different/no_cot
On disagreement the CoT answer becomes answer (5.2% of QAs). Nothing is
discarded, so you can filter to never-disagreed QAs or revert wholesale.
Style taxonomy (arity-split, per-style quotas at generation)
Single-doc: local_fact, calculation, multi_hop, multiple_choice.
Cross-doc: cross_revision_fact, cross_dependency_fact, cross_revision_mc,
cross_dependency_mc, cross_multi_hop. No cross-doc calculation.
doc_gap = j - i on cross-doc QAs is the long-range axis (1-7 for W=8).
Caveats
- Multiple-choice options are NOT in the
questionfield here — they live inchoices. The trainer's flattening renders them into the question, because closed-book eval shows the model the question and nothing else. - MC scores far higher than free-form (0.78 vs ~0.21 in our first run): five given options put the content in the prompt, so it measures discrimination, not recall from weights. Report it separately.
- The
summaryQA answer is 3-5 sentences, not a short span. It is training-only. - Cross-doc QAs are biased toward short answers: the generation parser rejects answers over 10 words, and that filter fires ~5x more often on cross-doc.
- Downloads last month
- 21