FineCat-NLI-Reason

A broad-coverage natural language inference training set (premise / hypothesis / label), with short reasoning traces and synthetic structural counterexamples. It extends FineCat-NLI with domain sources (clinical, scientific, legal, financial, fact-checking, RAG faithfulness, dialogue) and robustness data (monotonicity, negation, veridicality).

Configs

config	rows	columns
fast (default)	2,050,311 train / 33,347 validation	premise, hypothesis, label, label_space, source
think	121,768	premise, hypothesis, label, source, trace
structure	18,343 train / 508 test	premise, hypothesis, label, construction, trigger, half, base_source
rft	36,724 train / 1,936 validation	prompt, completion, label, source, kind, round, sampled_by, ...
dpo	18,948 train / 971 validation	prompt, chosen, rejected, label, source, ...
distill_fast	2,045,406 train / 33,225 validation	premise, hypothesis, label, label_space, source, teacher_logits
distill_think	222,766 train / 6,952 validation	premise, hypothesis, label, source, trace, round, ...

Training pipeline of the released two-mode model: SFT on fast + think → rejection-sampling fine-tuning of think mode on rft (2 rounds) → iterative RPO (DPO + NLL, 2 rounds) on dpo. Each section below describes exactly the rows used.

Distillation (4B student, FineCat-NLI v2 encoders): fast mode from distill_fast (teacher logits + gold), think mode of the 4B student from distill_think (the teacher's own correct traces).

Labels: 0 entailment, 1 neutral, 2 contradiction, 3 not-entailment (2-way sources, label_space = "2way", where neutral and contradiction are not distinguished).

fast additions: data/fast/train-anli_extra.parquet holds 34,963 ANLI train rows (rounds 1-3, gold labels, source = anli_extra) that FineCat-NLI had dropped (mostly easy items); pairs already in the mix or in ANLI dev / test were removed. They were added for the FineCat-NLI v2 encoders and the 4B student.

think: one short plain-prose reasoning trace per pair, generated by DeepSeek-V4.1-flash with a prompt asking it to reason only as far as needed. Only traces whose conclusion matches the gold label are kept. Traces contain no label line.

structure: synthetic pairs whose hypothesis reuses (almost) only the premise's words, built from real text in the fast sources, balanced between entailed and not-entailed:

• clauses under non-factive / conditional / modal wrappers vs factive / certain ones
• attachment misreads and role swaps vs modifier / conjunct drops and correct passives
• monotonicity: dropping or adding a modifier under downward (no, not, never, without, the restrictor of every) vs upward (some, several, a) contexts
• subsequences and constituents: the hypothesis is a contiguous span of the premise that is not entailed (garden paths such as "While the crew painted the hallway flooded." -> "The crew painted the hallway.", embedded subjects "suspected the invoices were forged", reduced relatives "The pharmacist trained at the clinic left", the consequent of a conditional, one disjunct) vs spans that are (the main clause after a factive because / after / although clause, the first clause of a subordinate construction, an object with its modifier dropped)

No HANS items were used to create them.

License

Rows come from existing datasets (see the source column); refer to each original source dataset for its license. Traces and synthetic pairs were generated with DeepSeek-V4.1-flash (the garden-path, embedded-subject and reduced-relative pairs with MiMo-V2.6-flash). Evaluation sets used for our benchmarks are excluded (exact and near-duplicate matches removed).

rft config (rejection-sampling fine-tuning data for think mode)

36,724 train / 1,936 validation reasoning traces in TRL's prompt-completion format (prompt, completion). prompt is the model's exact input (Premise: ... / Hypothesis: ..., no chat template); completion is a complete think block (<think>\n ... </think>); prompt + completion is exactly the text the model was trained on. These are the training rows of expert-iteration round 2 (ReST-EM style: each round samples from the previous round's model and retrains from the SFT checkpoint on all rounds so far, rounds 1, 2) for the think mode of a Qwen3.8-27B two-mode NLI model. The model trained on them is the starting point (and reference) of the dpo config's preference stage.

1. Sampling: per round 30,000 fresh training-split prompts (never reused across rounds), mostly those the current model is unsure about (P(gold) 0.2-0.95 without a trace), 6 traces each (temperature 1.0, top-k 20, top-p 0.95). Round 1 samples the SFT model (max 512 trace tokens), round 2 the round-1 model (max 1024) (sampled_by).
2. Correctness: a trace is correct when the model's classification head, reading the trace, predicts the gold label (two-way sources: not-entailment = neutral or contradiction). kind = mixed (the prompt had right and wrong samples; up to 2 distinct correct traces kept) or allcorrect (30% of such prompts, 1 trace).
3. Selection: among a prompt's distinct correct traces (up to 1024 tokens), shortest first (rank-based length preference; later replaced by random selection, which kept multi-step logic traces longer).
4. Label hygiene: round 1 includes noisy / convention-heavy sources whose gold was confirmed by a blind teacher (MiMo-v2.6-flash, never shown the label; disagreements dropped); round 2+ uses only sources with mechanically checkable gold. Entailment prompts that depend on an unnamed referent were dropped (strict reading). Round 1's VitaminC rows are excluded (built before the page titles were restored).

Columns: prompt, completion, label (0-3), label_name, source, tier, kind, round, sampled_by, pgold (head P(gold) after the trace), n_tokens (trace + close), prompt_id, sample_id.

Composition: {"mixed": 29488, "allcorrect": 9172} by kind; {"rft_round1": 23952, "sft": 14708} by sampler; {"neutral": 14181, "entailment": 12590, "contradiction": 8005, "not-entailment": 3884} by label.

dpo config (on-policy preference pairs for think mode: iterative RPO, 2 rounds)

18,948 train / 971 validation preference pairs in TRL's standard format (prompt, chosen, rejected). prompt is the model's exact input (Premise: ... / Hypothesis: ..., no chat template); chosen / rejected are complete think blocks (<think>\n ... </think>); prompt + chosen is exactly the text the model was trained on.

These pairs trained the released think mode with iterative Reasoning Preference Optimization (Pang et al. 2024), DPO (sigmoid, beta 0.1) plus an NLL term on the chosen trace. Each round starts from the previous round's model, which is also its DPO reference and the policy that sampled its pairs (rpo_round, sampled_by):

1. Round 1 (12,614 pairs): from the round-2 RFT model (trained on the rft config); NLL weight 1.0, lr 5e-6.
2. Round 2 (7,305 pairs): from the round-1 RPO model, on 30,000 fresh prompts plus 6,622 prompts that every earlier sample had answered wrong (prompt_origin = recovery); anti-displacement settings (NLL weight 1.5, DPO-Positive lambda 5, lr 3e-6). The released think mode is this run's step-200 checkpoint (the full run lengthened traces).

Pairs: per prompt with some samples right and some wrong (6 traces each, temperature 1.0, top-k 20, top-p 0.95, max 1024 trace tokens, each read by the model's own head): chosen = a random distinct correct trace with head P(gold) >= 0.6, rejected = the wrong trace closest in length to it with P(gold) <= 0.4 (length-matched); up to 2 pairs per prompt, no trace reused. Both sides are the policy's own text. Only sources with mechanically checkable gold.

Columns: prompt, chosen, rejected, label (0-3), label_name, source, tier, rejected_label, chosen_pgold, rejected_pgold, chosen_ref_logp, rejected_ref_logp (the round's reference = its sampling policy), chosen_n_tokens, rejected_n_tokens, rpo_round, sampled_by, prompt_origin, prompt_id, chosen_sample_id, rejected_sample_id.

Composition: {"contradiction": 7252, "neutral": 6474, "entailment": 5893, "not-entailment": 300} by label; {"fresh": 16669, "recovery": 3250} by prompt origin.

Earlier versions of this config held round-1 DPO pairs (with teacher-written chosen traces) from a run that was not used; they remain in the repository history.

distill_fast config (fast rows with teacher logits)

2,045,406 train / 33,225 validation: the fast rows (train includes anli_extra, not the train-subseq structural rows) with teacher_logits, the 27B teacher's fast-mode logits (the released two-mode model's SFT fast mode; no trace) in the order [entailment, neutral, contradiction], always 3-way: for 2-way sources (label_space = 2way, gold 0 or 3) not-entailment = neutral + contradiction after softmax. The FineCat-NLI v2 encoders and the 4B student's fast mode were trained on these rows with a KL term to the teacher plus cross-entropy on the gold label (the student at temperature 2). validation is the mix's held-out split (never trained on).

Columns: premise, hypothesis, label (0-3), label_space, source, teacher_logits (float32 x 3).

distill_think config (think traces used to distill a 4B student)

222,766 train / 6,952 validation reasoning traces written by the 27B teacher's think mode (Qwen3.8-27B two-mode NLI model, the line trained on rft and dpo) and used to train the think mode of a 4B student (Qwen3.5-4B, one LoRA adapter with a fast head and a think head). Same plain format as think: premise, hypothesis, label, trace (no think tags, no label line).

Two rounds (round), both sampled at temperature 1.0, top-k 20, top-p 0.95, max 1024 trace tokens, and judged by the teacher's own classification head reading the trace (correct = it predicts the gold label; two-way sources: not-entailment = neutral or contradiction):

1. r4 (62,335 rows; student steps 0-12,000): the prompts of the dpo config's round 2 (mechanically checkable sources, plus the recovery prompts), 6 samples each from the round-1 RPO model (sampled_by = rpo_round1); up to 2 distinct correct traces per prompt with head P(gold) >= 0.6, chosen at random.
2. r5 (167,383 rows; from step 12,000 on): about 190,000 fresh training-split prompts across all fast sources (equal quota per source, 40% of them ones the teacher's fast mode finds hard: P(gold) 0.2-0.95 or wrong), never used in an earlier round, 2 samples each from the released think mode (rpo_round2_step200). One trace per prompt where both samples were correct (kind = allcorrect), the correct one where only one was (mixed). Label hygiene: on noisy-tier sources the teacher's blind fast-mode answer had to agree with gold; SciTail / SRA entailment prompts dropped (strict reading); prompts whose premise had to be truncated to fit 2,048 tokens dropped.

r4 traces had reached ~2.7 passes each by student step 12,000, where held-out think quality began to fall; the run was rewound there and continued on r5 (10% of training tokens, about one pass per trace). validation = the pairs the student holds out for its think evaluation (never trained on).

Columns: premise, hypothesis, label (0-3), label_name, source, trace, round, sampled_by, kind (r5 only), pgold (teacher head P(gold) after the trace), n_tokens (trace tokens), pair_id, sample_id.

Composition: {"r4": 62335, "r5": 167383} by round; {"entailment": 92664, "neutral": 53522, "not-entailment": 44061, "contradiction": 39471} by label; median trace 53 tokens.