Submitted system (seed 42 of 3). Trained on 1,697 instances including 600 style-matched synthetic examples generated to match the target task's input distribution (~59-character responses, balanced classes), which is what lifted mean single-seed agreement from 0.67 to 0.84. Ensemble the three seeds and decode under the stated 10-per-class prior to reproduce the submission.

Predicts the degree of commitment a Japanese company's IR response expresses toward a future action, decision, or policy stance, on the five-level JF-ICR scale from the EBISU benchmark (arXiv:2602.01479):

Label	Meaning
`+2`	Strong Commitment — explicit, decided, unconditional
`+1`	Weak or Qualified Commitment — directional intent, hedged
`0`	Neutral or Hedged Intent — explanation without a forward stance
`-1`	Weak Refusal — time-bound or conditional rejection
`-2`	Strong Refusal — categorical, no room to reconsider

This is commitment strength, not sentiment: a warm, detailed answer that commits to nothing is 0.

Usage

python

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_id = "paragdakle/jficr-modernbert-ja-310m-v4-seed42"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id).eval()

text = "質問: 来期は増配を実施しますか。\n回答: 取締役会で1株当たり60円への増配を決議しました。"
with torch.no_grad():
    logits = model(**tokenizer(text, return_tensors="pt", truncation=True,
                               max_length=1024)).logits
print(model.config.id2label[logits.argmax(-1).item()])  # -> "+2"

Input format matters — the model was trained on 質問: {question}\n回答: {response}.

Training data

1,697 instances: 253 public JF-ICR + 477 LLM-consensus (silver) instances harvested from real Japanese IR transcripts + 376 verified minimal pairs (265 refusal-side, 111 +2) + 600 style-matched synthetic instances (120 per class).

Silver labels come from a three-judge LLM ensemble (claude-opus-4-8, gpt-5.5, gpt-5.6-sol) with tiered consensus (unanimous / majority / ordinal-median) and a spread gate. Minimal pairs are real responses rewritten by claude-opus-4-8 to change only the stance-bearing sentence-final clauses, with the label shifted to match and confirmed by blind GPT judges (gpt-5.5, gpt-5.6-sol) that never saw the intended label — deliberately a different model family from the generator, so verification is independent. Topic, entities, figures, and register are held fixed across a pair.

Results

FinNLP 2026 public test slice (15 items, as an ensemble): accuracy 1.00
our reading of all 50 test items (3-seed ensemble, balanced decoding): 45/50 = 0.90
our reading of all 50 test items (this seed alone, argmax): 44/50

Limitations

Silver and synthetic data dominate. Roughly two thirds of the training pool is LLM-labeled or LLM-generated. Labels carry that noise; the paper's best reported model reaches only ~0.61 accuracy on this task, so consensus labels are wrong a meaningful fraction of the time.
Evaluate against your own distribution. The two checkpoints rank in opposite order on a naturally-skewed transcript split versus the short, clean, balanced competition items. Neither split predicts the other.
Small evaluation sets. Every number above rests on 50-51 items, with some classes at supports of 1-6. Differences of a few points are noise.
Narrow provenance. Source transcripts are Japanese mega-cap financials and trading houses (2022-2026). Mid-caps, other sectors, and other market regimes are unrepresented.
Not suitable for automated investment decisions or regulatory interpretation without human review.

Citation

The task and label scheme are from EBISU:

bibtex

@article{ebisu2026,
  title={Ebisu: Benchmarking Large Language Models in Japanese Finance},
  journal={arXiv preprint arXiv:2602.01479},
  year={2026}
}