Athekunal/llm-jev-reason-long

LLM-as-Jev Reason+Long mix, with generated evidence spans

The training and evaluation data used to train llm-jev-qwen3-0.6b and llm-jev-qwen3-4b, with its sources, pinned revisions and licenses listed below and, for a part of the training rows, evidence spans proposed by an LLM and scored with a SelfCite-style reward.

Licenses differ by source and some are non-commercial. This repository has no single license. Read the table below and each source's terms before any use beyond research.

What is in it

The mix has 13 datasets with 1,000 training rows each (13,000). ReClor is withheld because its data is research-only and not for redistribution, so this release has 12,000 training rows and 5,815 evaluation rows. The models were trained on the full 13,000 rows including ReClor.

config	file	rows	what
default	data/train.parquet	12,000	training decisions
default	data/eval.parquet	5,815	held-out decisions, disjoint from training by document group
spans	spans/train.parquet	6,781	generated evidence for training rows of ProofWriter, QuALITY, alpha-NLI, ContractNLI, CosmosQA, MuSiQue, LogiQA 2.0 and Nimble

Decision rows (data/*.parquet)

column	meaning
id	<source>:<hash>, unique
source	dataset name
parts	the input text, in prompt order: list of {name, text}
question	the question
options	ordered options; option i is shown as number i + 1
gold	0-based index of the correct option
source_spans	evidence that the source dataset itself provides, not generated by us: list of {part, start, end} character ranges into parts[part].text (ProofWriter, ContractNLI, MuSiQue, Nimble; empty elsewhere). How each source made it differs (annotators, proofs, verified pairs); see the source. The LLM-generated evidence is in the spans config
group	leakage unit (document, story, ...); one group stays in one split
subset	stratification key
meta	JSON string of extra provenance

Eval and training rows were checked against the external benchmarks used in the paper (13-gram overlap) and overlapping rows were removed.

Generated evidence (spans/train.parquet)

A labeller model (Qwen3.6-27B with thinking, served with vLLM) read each document split into sentences and, in 8 independent samples, cited the smallest set of sentences that supports the gold answer. Distinct cited sets are scored, following SelfCite (Chuang et al., arXiv 2502.09604):

• necessity = log q(gold | x) − log q(gold | x without S)
• sufficiency = log q(gold | S only) − log q(gold | x)
• reward = sufficiency + necessity = log q(gold | S only) − log q(gold | x without S)

where q is the labeller's own probability of the gold option. The set with the highest reward becomes the row's evidence spans.

column	meaning
id, source	as in the decision rows
log_q_gold	log q(gold | full input)
model_answer, gold	the labeller's own answer and the gold option; the reward is uninformative where they disagree
n_parsed	how many of the 8 samples gave a valid citation block
cut_off	samples cut off by the token cap
completion_tokens	mean tokens per sample
candidates	the distinct cited sets: sentences (0-based sentence indices of the row), votes (how many of the 8 samples proposed it), necessity, sufficiency, reward
selected	index into candidates of the highest-reward set
spans	the selected set as character ranges {part, start, end} into parts
span_text	the text of those ranges

Only the parsed outputs of the 8 samples are released as candidates with their votes; the raw reasoning text of the samples was not kept. Sampling settings are those of llm_jev.span_generation.best_of_n and are not stored in the files. Sentence indices follow llm_jev.modeling.context_cite.split_sources. The row's spans are the unambiguous way to use the evidence.

Code that builds the mix and the spans: https://github.com/Athe-kunal/llm-as-jev