Datasets:

idealab-cs2
/

italic-extkd-pool

license: apache-2.0
language:
  - it
task_categories:
  - question-answering
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.jsonl
      - split: italic_sft
        path: italic_sft.jsonl
      - split: italic_sft_ext
        path: italic_sft_ext.jsonl
      - split: quiz_militare
        path: quiz_militare.jsonl
      - split: pinocchio
        path: pinocchio.jsonl
tags:
  - italian
  - multiple-choice
  - knowledge-distillation
  - italic

italic-extkd-pool

The stage-3 training data of idealab-cs2/zagreus-0.4B-italic-extkd: 57,563 Italian multiple-choice questions from public, in-distribution datasets with teacher soft labels. One soft-KD stage from the stage-2 checkpoint on the agreement-filtered subset (28,561 items where the teacher agrees with the gold answer) reaches 0.4921 / 0.4929 / 0.4932 on the full ITALIC 10K (official harness, 5-shot fast, temperature 0, three independent runs). Full lineage: 0.2802 base -> 0.4787 stage 1 -> 0.4878/0.4880 stage 2 -> this stage.

train is the full pool; the other four splits partition it by provenance:

split	rows	contents
`train`	57,563	the full labeled pool (union of the four below)
`italic_sft`	20,665	FinancialSupport/italic_sft: ITALIC-style instruction items from the challenge ecosystem's public data pipeline, 13 category source-splits
`italic_sft_ext`	3,268	FinancialSupport/italic_sft_ext: extended ITALIC-style items
`quiz_militare`	3,641	FinancialSupport/quiz_militare: Italian public-exam knowledge MCQs
`pinocchio`	29,989	efederici/pinocchio sample: general/culture/law Italian exam questions

Each row: question, options (lettered), answer (gold letter), category, source (provenance), index, teacher_logprobs (soft labels: per-letter log-probabilities from Mistral-Small-3.2-24B-Instruct-2506, public and run locally, prompted in the ITALIC format), teacher_letter, teacher_correct (teacher/gold agreement flag — the model trained on the teacher_correct == true subset of 28,561 items).

Composition note: ~14K questions are shared with the stage-1/stage-2 pools (rehearsal); ~40K are new at this stage, led by the italic_sft/quiz_militare items. The pools are cumulative by design, not independent.

Decontaminated against the ITALIC test set: exact match plus semantic similarity at cosine 0.80 — 0 exact leaks and 0 near-duplicates for italic_sft / italic_sft_ext / quiz_militare; the pinocchio sample showed 0.74% near-duplicates which were dropped (manifest in the submission repo). All splits are answer-position balanced (~25% per letter A–D): position skew in synthetic MCQ pools was found to cause answer-distribution collapse in the student, so balance is verified and documented.

Part of the submission for the mii-llm Italian Post-Training Challenge.

Model trained on this data: idealab-cs2/zagreus-0.4B-italic-extkd
Companion pools: italic-softkd-pool (stage 1) · italic-m2-culture-pool (stage 2)
Training code and report: github.com/ruggsea/italic-challenge-submission

from datasets import load_dataset
ds = load_dataset("idealab-cs2/italic-extkd-pool")               # all five splits
core = load_dataset("idealab-cs2/italic-extkd-pool", split="italic_sft")
train = [r for r in ds["train"] if r["teacher_correct"]]          # the 28,561-item training subset

Training, data curation and evaluation by ruggsea, using compute resources of the Complex Social & Computational Systems (CS²) group, IDea_Lab, University of Graz.

License: Apache 2.0.