calibrated-typed-decisions (v1)

Typed decisions (multiple choice and yes/no) over short English texts, each with a human gold label and a calibrated soft label from a teacher LLM. Built for training and evaluating small decision models whose confidence scores mean something.

Why

A model that says 0.9 should be right about 90% of the time. Instruction-tuned LLMs read as classifiers tend to be overconfident. This dataset pairs every decision with a teacher probability distribution that has been calibrated against real answers, plus the raw ingredients so you can redo the calibration your own way.

Splits

split	rows
train	4998
validation	1500
test	1499

Fields

field	meaning
type	choice or noul (yes/no; options are always ["false", "true"])
instructions	the question
state	the text being judged
options	option labels in canonical order
gold, gold_index	the human label from the source dataset
teacher_option_logprobs	teacher log-probability of each option's answer letter, canonical order
teacher_temperature	temperature fit for this row's (source, type) group
teacher_shift	yes/no rows only: offset added to the true score after dividing by the temperature (0 for choice rows)
teacher_probs	softmax(teacher_option_logprobs / teacher_temperature + [0, teacher_shift]) for yes/no, without the shift for choice
teacher_letter_mass	total probability the teacher put on answer letters at all
display_perm	the option order the teacher saw (display position to canonical index)

How the labels were made

1. Examples come from public, permissively licensed classification datasets (listed below).
2. Each example becomes a question with 3 to 8 shuffled options (the gold label plus random distractors), or a yes/no question about one candidate label (about 30% of rows).
3. The teacher (Qwen/Qwen2.5-7B-Instruct, 4-bit NF4) reads each question once. The answer-letter log-probabilities at the first answer position are recorded. Nothing is generated.
4. Rows where the teacher put under 50% of its probability on answer letters are dropped (train 2, validation 0, test 1).
5. One temperature per (source, type) is fit on the validation split by minimizing log loss against gold labels, then applied to every row. Yes/no groups also fit a shift toward true or false (a two-parameter Platt-style fit), because the teacher leaned toward answering "no": a temperature alone cannot correct a lean. Groups with fewer than 30 validation rows borrow a pooled fit from all sources of the same type (fit = pooled). Calibration is reported on the test split.

Teacher calibration on the test split

Accuracy can change after calibration for yes/no rows, because the shift can flip a borderline answer. Temperature alone never changes the top pick. "ECE dial only" shows the temperature without the shift, for comparison.

source	type	n	acc raw	acc calibrated	T	shift	fit	ECE raw	ECE dial only	ECE calibrated
clinc/clinc_oos	choice	209	0.957	0.957	3.67	+0.00	own	0.046	0.027	0.027
clinc/clinc_oos	noul	91	0.824	0.901	1.36	+6.95	own	0.183	0.129	0.067
google-research-datasets/go_emotions	choice	216	0.602	0.602	8.29	+0.00	own	0.360	0.064	0.064
google-research-datasets/go_emotions	noul	84	0.738	0.810	9.89	+0.75	own	0.245	0.110	0.119
mteb/amazon_massive_intent	choice	214	0.921	0.921	4.57	+0.00	own	0.077	0.024	0.024
mteb/amazon_massive_intent	noul	86	0.814	0.895	3.51	+2.15	own	0.180	0.077	0.091
mteb/amazon_polarity	choice	216	0.954	0.954	3.92	+0.00	own	0.045	0.024	0.024
mteb/amazon_polarity	noul	84	0.952	0.952	3.51	+0.00	own	0.041	0.055	0.055
mteb/banking77	choice	205	0.946	0.946	4.28	+0.00	own	0.052	0.025	0.025
mteb/banking77	noul	94	0.798	0.830	5.22	+1.05	own	0.165	0.064	0.095

Sources

• CLINC150 (Larson et al., 2019), CC-BY-3.0
• GoEmotions (Demszky et al., 2020), Apache-2.0
• BANKING77 (Casanueva et al., 2020), PolyAI/banking77, CC-BY-4.0; loaded via mteb/banking77
• MASSIVE (FitzGerald et al., 2022), English subset, CC-BY-4.0; loaded via mteb/amazon_massive_intent
• Amazon Review Polarity (Zhang et al., 2015), Apache-2.0 per fancyzhx/amazon_polarity; loaded via mteb/amazon_polarity

Limitations

• Soft labels reflect one teacher's judgment and biases, not human disagreement.
• Calibration is fit per group on average; individual rows can still be off.
• Distractors are sampled at random, so many choice questions are easier than hard-negative ones.
• Options are humanized source label names without descriptions.
• English only. Gold labels inherit any noise in the source datasets.
• A toxicity source (Civil Comments) was scored and left out: the teacher agreed with its crowd labels at roughly chance level, so its soft labels carried no reliable signal.
• Yes/no groups: the shift improved accuracy and corrected the teacher's lean toward "no", but in some groups the temperature alone scored slightly better on ECE. Both are reported.
• The teacher ran in 4-bit; full-precision logits would differ slightly.

Usage

from datasets import load_dataset
ds = load_dataset("m-lagnajit/calibrated-typed-decisions")
row = ds["train"][0]
print(row["options"], row["teacher_probs"], row["gold"])