U-WIN/equational-theory-sair-notes

Every notes row now includes original_question_text (list<string>). The texts are copied verbatim from the original source questions/tasks; there are no null, blank, synthesized or truncated replacements. Bank lists follow related_question_ids in the same order. Initial Equational R0 notes without a question ID contain the exact assigned task recovered from their saved source-session transcript. In rollout notes tables, the list contains the exact original task for that row’s task_id. Every prior column, record, ordering and training text is unchanged. Both JSONL and Parquet include the new column. Current export hashes and exhaustive coverage checks are in provenance/question-text-column.json; earlier export hashes and byte-prefix statements describe the preceding release.

Equational-theory SAIR notes

Completed-round update prepared 2026-10-05T22:57:58.353088+00:00.

Configuration	Continuation generator	Completed through	Usable notes	Training-format note tokens
notes_recursive	Qwen/Qwen3.5-9B	R13	545,275	324,831,419
notes_recursive_checkpoint	violetxi/qwen35-9b-equational-theory-sair-mix100m-70n30t-thinking	R14	511,571	299,995,882

The base and checkpoint banks are separate continuations of the same previously published baseline. The listed generator applies to the continuation rounds; inherited baseline notes retain their original provenance.

banks/notes_recursive.* selects the base continuation. branches/checkpoint/banks/notes_recursive.* selects the checkpoint continuation. Both contain their full cumulative live bank. Existing dataset configurations and historical raw archives remain available.

Only finalized, merged rounds are included. Active worker checkpoints, unfinished rounds, and rejected/quarantined notes are excluded. Per-branch provenance/source-snapshot.json, provenance/completed-rounds.json, and validation.json record the source boundary and exact file validation.

Parquet field names and types match the previous publication. C&H JSONL preserves its original bytes; the existing Parquet-only written_by annotation is retained and extended from finalized agent/connective outputs.

Earlier release and protocol documentation

Equational-theory SAIR notes

192,372 current usable notes, with 113,692,388 title/body tokens, from the completed R0–R8 bank. Updated October 1, 2026. The training rendering with two trailing newlines contains 113,810,206 tokens. This uses the JSONL/Parquet layout and original field names/types of violetxi/harvey-notes-v4. Only a recursive bank exists for this experiment; there is no inherited-only comparison bank.

Round of current revision	Notes
R0	4,870
R1	24,666
R2	26,228
R3	26,148
R4	21,364
R5	21,261
R6	22,611
R7	22,667
R8	22,557

Files and schema

• banks/notes_recursive.jsonl and .parquet: identical rows in identical order, sorted by round, revision creation time, and note ID. The Parquet copy is browsable with the dataset viewer.
• The original Harvey fields are preserved: note_type, title, text, written_by, generation_round, generation_batch, evidence, supporting_trajectory_ids, related_question_ids, prompts_version, note_id, repository, commit, and training_text.
• Added mathematical provenance: revision, source_session_id, equation_ids, claims_json, and open_questions_json. Supporting trajectory IDs are note-generation session IDs. Related question IDs are admitted proposals from that source session. written_by=agent; prompts_version is null, following the reference's agent-note convention. Frozen prompt files are in procedure/.
• raw/: historical R0–R6 archive of note revisions, quarantined heads, collection sessions, questions, problems, attempts, rounds and initial tasks. These files are unchanged in this notes-only update; their scope is recorded in provenance/raw-source-snapshot.json. The current live notes bank and current provenance cover completed R0–R8.
• banks/notes_recursive.stats.json, provenance/, and validation.json: current counts, completed-round snapshot and merge receipts, and full JSONL/Parquet equality checks. The training-note-snapshot receipt describes the historical R0–R4 training input.

Content and training use

These are Qwen3.5-9B self-study notes about magmas, generated from the SAIR study projection of teorth/equational_theories at 98a39927ae9ab8ee58d37587231df80563849160. Note types are statement, derivation, counterexample, and technique. Model revision: c202236235762e1c871ad0ccb60c8ee5ba337b9a.

training_text is title + newline + body, preserved exactly from the equational experiment. It does not add the Harvey repository header. Every published note ID, revision and training text was checked against the immutable completed R8 snapshot. Content-token counts use the pinned Qwen tokenizer without special tokens or truncation. Effective training-token counts depend on downstream packing and loss masking and are not reported for this update. The original R0–R4 dataset used by the frozen 100M mixture remains accessible at revision fbc728524ae3a2c8cc6a8806830698d3878fac7c; provenance/training-note-snapshot.json and provenance/history/R0-R4-training-token-counts.json describe that historical input.

The bank excludes 4,084 quarantined heads. No additional quality filter is applied. This update adds 45,224 note IDs relative to the prior R6 HF revision; all 147,148 prior usable rows are preserved exactly. These are current revision heads, not an append-only history. R7 and R8 are complete and merged; unmerged worker checkpoints are never mixed into this snapshot. The existing raw/ history remains the historical R0–R6 archive. The R6 live notes bank remains available at revision ef723fb22b14da0014a546c27180604c31efa7a5.

SAIR task eligibility is governed by the study holdout policy. Notes are model-generated mathematical claims, not certified Lean proofs; generation gates do not establish every note's correctness.

from datasets import load_dataset
notes = load_dataset("violetxi/equational-theory-sair-notes", "notes_recursive", split="train")