wiiiktor/chopper_demo

Chopper data

70-second video: youtu.be/KMvmdXCEOJU

Code: github.com/wiiiktor/chopper_demo (scripts/data/).

Look at the data first

examples.md shows four rows joined across the files (MuSiQue, MuSiQue with a missing hop, HotpotQA, TriviaQA): question, gold answer, the kept fragments, the teacher's reasoning steps, the ANSWER/SEARCH decision and the follow-up question. examples.jsonl has the same rows with all 16 passages. One of them:

Question: In 2012, how many non-Hispanic Whites lived in the city where the creator of La Grande Vitesse died? Gold answer: 2.7 million · source: dataset: "musique", row: 17422

16 passages, 12,924 characters → 3 fragments, 362 characters (2.8%):

• La Grande Vitesse: "La Grande Vitesse, a public sculpture by American artist Alexander Calder"
• New York City: "New York City has the largest European and non-Hispanic white population of any American city."
• New York City: "At 2.7 million in 2012, New York's non-Hispanic white population is larger than …"

The passage saying that Calder died in New York is not among the 16, so the decision is SEARCH, with the follow-up question "Where did Alexander Calder die?"

Which benchmark each row comes from

Every row of every file carries "dataset": "musique", "hotpot" (HotpotQA) or "triviaqa", plus "row", the index in that benchmark's train split (validation rows: the val split's index). dataset + row is the key that joins pools_, ctx_, qw_ and subq_. Dataset B uses "wiki2hop" / "wiki1hop".

split	MuSiQue	HotpotQA	TriviaQA
train	1,200	800	1,000
train2	2,400	1,600	2,000
train3	6,000	6,000	6,000
val	100	100	100

benchmark_chains/ (dataset A)

27,000 training questions from the train splits of MuSiQue (9,600), HotpotQA (8,400) and TriviaQA (9,000), plus 300 validation questions (100 each). Do not report models trained on it as zero-shot on these benchmarks.

files	what
pools_<split>.jsonl + chunks_<split>.json	1,000-chunk pools per question; pools reference the shared chunk table by index
ctx_<split>.jsonl	top 16 after Qwen3-Embedding-0.6B (top 100) → bge-reranker-v2-m3
qw_<split>.jsonl	Qwen3-14B-AWQ fragment labels as character spans [passage, start, end] (the teacher saw the gold answer)
subq_<split>.jsonl	follow-up questions for rows with a missing hop

Splits: train (3,000), train2 (6,000), train3 (18,000), val (300).

wikipedia_chains/ (dataset B)

Benchmark-free: questions written by Qwen3-14B-AWQ from English Wikipedia (wikimedia/wikipedia 20231101.en) and filtered by Qwen3-4B. 2-hop ("dataset": "wiki2hop"): article A mentions article B by its title, the answer is a short fact in B, and the question describes B only through A. 1-hop ("wiki1hop"): one paragraph, a question that stands on its own. "row" indexes the split. Same file layout as dataset A.

split	wiki2hop	wiki1hop	14B fragment labels	follow-up questions
wtrain	18,000	9,000	27,000	1,164
wval	200	100	300	—

Acceptance by the Qwen3-4B filter: ~40.5% of 2-hop and ~80% of 1-hop questions; Qwen3-4B answers only 7–9% of the accepted ones without a passage. Pools: the gold paragraph(s), the other paragraphs of A and B and of up to 8 articles A mentions (hard negatives), then random article leads, 1,000 chunks per question. Train and val share no A or B article.

Licenses follow the sources: MuSiQue CC BY 4.0, HotpotQA and Wikipedia CC BY-SA 4.0, TriviaQA Apache 2.0 ⇒ CC BY-SA 4.0.