dleemiller / finecat-nli-v2-m

FineCat-NLI v2 Medium

Natural Language Inference (NLI): the Swiss Army knife of encoders

Provide a text (premise) and statement (hypothesis), and it returns calibrated probabilities that the text entails, contradicts, or is neutral to the statement. A multiple-choice question is one statement per option.

Examples

150M parameters · documents up to 8k tokens · ~1532 pairs/s · Apache-2.0

from sentence_transformers import CrossEncoder

model = CrossEncoder("dleemiller/finecat-nli-v2-m")
text = "My card was charged twice for the same order."
labels = ["billing", "shipping", "account access"]
probs = model.predict([(text, f"This text is about {l}.") for l in labels], apply_softmax=True)
print({l: round(float(p[0]), 3) for l, p in zip(labels, probs)})  # p[0] = P(entailment)
# {'billing': 0.62, 'shipping': 0.16, 'account access': 0.03}

More in Recommended Applications and the 45-dataset evaluation .

---

Table of Contents

• NLI Evaluation Results
• Usage
• Recommended Applications
• Extended Evaluation (45 datasets)
• Training Details
• Acknowledgments
• Citation

---

NLI Evaluation Results

F1-Micro scores (equivalent to accuracy) for each dataset. Performance was measured at bs=32 using a Nvidia Blackwell PRO 6000 Max-Q.

45-set avg: macro accuracy over the 45 datasets , each model at its own maximum context (512 tokens for the DeBERTa and RoBERTa models).

---

Usage

Label Map:

• entailment : 0
• neutral : 1
• contradiction : 2

Direct Usage (Sentence Transformers)

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load the model and run inference.

from sentence_transformers import CrossEncoder
import numpy as np

model = CrossEncoder("dleemiller/finecat-nli-v2-m")
id2label = model.model.config.id2label  # {0:'entailment', 1:'neutral', 2:'contradiction'}

pairs = [
    ("The glass fell off the counter and shattered on the tile.",
     "The glass broke when it hit the floor."),          # E
    ("The store opens at 9 a.m. every day.",
     "The store opens at 7 a.m. on weekdays."),          # C
    ("A researcher presented results at the conference.",
     "The presentation won the best paper award."),       # N
    ("It started raining heavily, so the match was postponed.",
     "The game was delayed due to weather."),             # E
    ("Every seat on the flight was taken.",
     "There were several empty seats on the plane."),     # C
]

logits = model.predict(pairs)  # shape: (5, 3)

for (prem, hyp), row in zip(pairs, logits):
    pred_idx = int(np.argmax(row))
    pred = id2label[pred_idx]
    print(f"[{pred}]  Premise: {prem}  |  Hypothesis: {hyp}")

---

Recommended Applications

Frame each check as a premise (text) and hypothesis (statement), and use the entailment probability.

Less suited: financial and general legal verification (findver, legal_domain), clinical-trial eligibility (nli4ct), multi-document legal reasoning (control), and adversarial reasoning (anli_r3, nope_adv).

---

Extended Evaluation (45 datasets)

Accuracy on the test splits of 45 NLI and adjacent datasets. 2-way datasets (entailed vs. not) are scored as entailment vs. any non-entailment prediction. DocNLI and VitaminC are evaluated on fixed random subsets of 1,000 and 10,000 pairs. FRANK (news-summary factuality) is scored per summary sentence: each sentence vs. its source article, labelled by the majority annotator judgment.

Macro average over the 45 datasets: 0.7967 .

Training Details

This model is a fine-tune of jhu-clsp/ettin-encoder-150m , trained on the dleemiller/FineCat-NLI-Reason dataset, a compilation of several high-quality NLI data sources with quality screening and reduction of easy samples in the training split. Also includes additional NLI, fact verification and long document sources (2.05M pairs in total). The training uses logit distillation from an NLI finetune of Qwen3.8-27B .

Distillation loss looks like this: $\begin{array}{cccc}& L=\alpha \cdot L_{\text{CE}}(z^{\left(s\right)},y)+\beta \cdot T^2\cdot KL\left(\sigma \right(z^{\left(t\right)}/T\left)\parallel \sigma \right(z^{\left(s\right)}/T\left)\right)& & \end{array}\backslash begin\{equation\}\; \backslash mathcal\{L\}\; =\; \backslash alpha\; \backslash cdot\; \backslash mathcal\{L\}\_\{\backslash text\{CE\}\}(z^\{(s)\},\; y)\; +\; \backslash beta\; \backslash cdot\; T^2\; \backslash cdot\; \backslash mathrm\{KL\}\backslash left(\backslash sigma(z^\{(t)\}/T)\; \backslash ,\backslash |\backslash ,\; \backslash sigma(z^\{(s)\}/T)\backslash right)\; \backslash end\{equation\}$ L = α ⋅ L CE ​ ( z ( s ) , y ) + β ⋅ T 2 ⋅ KL ( σ ( z ( t ) / T ) ∥ σ ( z ( s ) / T ) ) ​ ​

where $z^{\left(s\right)}z^\{(s)\}$ z ( s ) and $z^{\left(t\right)}z^\{(t)\}$ z ( t ) are the student and teacher logits, $yy$ y are the ground truth labels, $\alpha =0.2\backslash alpha\; =\; 0.2$ α = 0.2 , $\beta =0.8\backslash beta\; =\; 0.8$ β = 0.8 and $T=1T\; =\; 1$ T = 1 .

Training inputs keep the full documents the teacher labelled (up to 8k tokens). The model was trained for 3 epochs at a constant learning rate followed by a 1-epoch cooldown. The released weights average the cooldown's final weights with the epoch-3 exponential-moving-average checkpoint (greedy soup), with the output temperature calibrated on the validation split. The learning rate was 5e-5.

---

Acknowledgments

We thank the creators and contributors of jhu-clsp (Ettin) and the Qwen team for making their work available. This model would not be possible without their efforts and open source contributions.

Citation

@misc{nli-compiled-2025,
  title = {FineCat NLI Dataset},
  author = {Lee Miller},
  year = {2025},
  howpublished = {Refined compilation of 6 major NLI datasets}
}