README.md

metadata

dataset_info:
  features:
    - name: id
      dtype: string
    - name: source
      dtype: string
    - name: source_license
      dtype: string
    - name: source_id
      dtype: string
    - name: type
      dtype: string
    - name: instructions
      dtype: string
    - name: state
      dtype: string
    - name: options
      list: string
    - name: option_descriptions
      list: string
    - name: candidate
      dtype: string
    - name: gold
      dtype: string
    - name: gold_index
      dtype: int64
    - name: display_perm
      list: int64
    - name: teacher_option_logprobs
      list: float64
    - name: teacher_temperature
      dtype: float64
    - name: teacher_shift
      dtype: float64
    - name: teacher_probs
      list: float64
    - name: teacher_letter_mass
      dtype: float64
    - name: human_probs
      list: float64
    - name: soft_label
      list: float64
    - name: soft_label_source
      dtype: string
    - name: corrected_soft_label
      list: float64
    - name: soft_label_corrected
      dtype: bool
  splits:
    - name: train
      num_bytes: 12924237
      num_examples: 16995
    - name: validation
      num_bytes: 1379863
      num_examples: 1799
    - name: test
      num_bytes: 1377093
      num_examples: 1797
    - name: heldout
      num_bytes: 1915788
      num_examples: 1000
  download_size: 11013876
  dataset_size: 17596981
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
      - split: heldout
        path: data/heldout-*
language:
  - en
license: cc-by-4.0
task_categories:
  - text-classification
pretty_name: Calibrated Typed Decisions
tags:
  - calibration
  - soft-labels
  - distillation
  - typed-decisions

calibrated-typed-decisions (v2.1)

Typed decisions (multiple choice and yes/no) over short English texts. Every row has a human gold label and a calibrated soft label, plus the raw ingredients to rebuild the soft label your own way. Earlier versions stay available: load_dataset("m-lagnajit/calibrated-typed-decisions", revision="v1.0") and revision="v2.0".

What changed in v2.1

• New column corrected_soft_label: identical to soft_label, except on rows where the calibrated teacher's top pick disagreed with the human answer. There, the teacher's value on its top pick and its value on the right answer are swapped: the teacher's numbers are kept, but the right answer gets the largest one. If the right answer is already tied for the teacher's largest value, the row is unchanged. Rater-labelled GoEmotions rows are unchanged; held-out rows stay empty.
• New column soft_label_corrected: true on exactly the rows where that swap changed the values.
• Nothing else changed. soft_label, teacher_probs and every other value are exactly as in v2.0.

How often the teacher's top pick was wrong (train, validation and test):

source	type	rows changed	teacher-labelled rows	share
amazon_massive_intent	choice	220	2516	8.7%
amazon_massive_intent	noul	108	1076	10.0%
amazon_polarity	choice	93	2496	3.7%
amazon_polarity	noul	36	1104	3.3%
banking77	choice	284	2484	11.4%
banking77	noul	146	1116	13.1%
clinc_oos	choice	123	2461	5.0%
clinc_oos	noul	104	1139	9.1%
scicite	choice	408	1832	22.3%
scicite	noul	205	768	26.7%

What changed from v1

• GoEmotions soft labels come from the human raters, not the teacher. On 600 GoEmotions validation and test questions, the teacher's top pick disagreed with the human answer 38.5% of the time, which is why GoEmotions soft labels in v2 come from the raters instead.
• Options carry short descriptions (card arrival: the customer is waiting for a card). Some are hand-written; the rest were written by the teacher from real training examples.
• Harder distractors: about half the wrong options are the labels closest to the right answer (by embedding similarity of label and description), so questions aren't trivially easy.
• More data: 3,000 training questions per source (2,000 for SciCite), plus SciCite as a new source.
• A held-out split from PubMedQA (see below).

Splits

split	rows	use
train	16995	training
validation	1799	choosing epochs, fitting dials
test	1797	final evaluation on the same sources
heldout	1000	test only: a new source and a new kind of decision

Do not train on heldout, and do not fit dials on it. It's 1,000 expert-labelled PubMedQA questions (yes / no / maybe from an abstract), there to measure how well a model decides on a label set and domain it has never seen.

Fields

field	meaning
type	choice or noul (yes/no; options are always ["false", "true"])
instructions, state	the question, and the text being judged
options, option_descriptions	option labels in canonical order, and their descriptions
candidate	yes/no rows: the label being asked about
gold, gold_index	the human label from the source dataset
soft_label	the recommended training target: rater votes for GoEmotions, the calibrated teacher otherwise
soft_label_source	human_raters, teacher, or none (held out)
corrected_soft_label	soft_label with the teacher's wrong top picks corrected (see v2.1 above)
soft_label_corrected	true where corrected_soft_label differs from soft_label
human_probs	GoEmotions only: rater votes per option, plus 0.1 pseudo-votes per option
teacher_option_logprobs	the teacher's raw log-probability of each option's answer letter
teacher_temperature, teacher_shift	the dial and nudge fit on validation for this row's group
teacher_probs	the calibrated teacher distribution
teacher_letter_mass	how much probability the teacher put on answer letters at all
display_perm	the option order the teacher saw
source_id	the row's id in its source dataset, where one exists

Which training target to use

We trained ModernBERT-base students on the train split with 6 different targets, each with 3 random seeds (3 epochs; best epoch chosen on validation log loss). Every student then gets 2 temperature dials fitted on validation: one for choice questions, and one plus a yes/no shift for yes/no questions. We compared this with one dial per source and question type, and with no dial, by cross-fitting on validation: 2 dials gave the lowest log loss for all five targets.

run	target
A	one-hot gold
B	gold, label-smoothed (0.1)
C	soft_label
D	0.5 * one_hot(gold) + 0.5 * soft_label
E	corrected_soft_label
F	0.5 * one_hot(gold) + 0.5 * corrected_soft_label

Single students: mean over 3 seeds (lowest–highest)

run	test accuracy	test ECE	test log loss	held-out accuracy	held-out ECE	held-out log loss
A	0.912 (0.909–0.917)	0.011 (0.006–0.019)	0.247 (0.230–0.267)	0.580 (0.538–0.614)	0.130 (0.097–0.171)	0.970 (0.945–0.997)
B	0.918 (0.913–0.923)	0.022 (0.016–0.026)	0.281 (0.268–0.289)	0.579 (0.568–0.585)	0.092 (0.020–0.149)	0.987 (0.957–1.005)
C	0.892 (0.886–0.898)	0.021 (0.020–0.023)	0.310 (0.307–0.313)	0.583 (0.576–0.592)	0.055 (0.027–0.106)	0.940 (0.926–0.952)
D	0.915 (0.909–0.918)	0.016 (0.015–0.019)	0.248 (0.241–0.256)	0.574 (0.548–0.595)	0.044 (0.028–0.064)	0.947 (0.878–1.017)
E	0.912 (0.908–0.915)	0.015 (0.014–0.016)	0.261 (0.252–0.267)	0.596 (0.591–0.605)	0.086 (0.048–0.107)	0.935 (0.924–0.946)
F	0.919 (0.917–0.921)	0.015 (0.015–0.016)	0.245 (0.241–0.249)	0.595 (0.584–0.607)	0.046 (0.019–0.068)	0.914 (0.888–0.953)

Ensembles: each run's 3 calibrated students averaged

run	test accuracy	test ECE	test log loss	held-out accuracy	held-out ECE	held-out log loss
A	0.922	0.018	0.221	0.595	0.153	0.964
B	0.925	0.009	0.249	0.582	0.052	0.948
C	0.895	0.013	0.292	0.595	0.065	0.932
D	0.920	0.008	0.231	0.589	0.054	0.928
E	0.918	0.013	0.241	0.587	0.081	0.926
F	0.928	0.006	0.225	0.595	0.051	0.899

On held-out, always answering "yes" scores 0.552 accuracy, and always predicting the label frequencies gives a log loss of 0.937.

What held up under both checks

Listed only if it held in every seed and in a paired bootstrap over questions for the ensembles.

Test split (the six training sources):

• C worse than A on accuracy
• C worse than A on log loss
• C worse than B on accuracy
• C worse than B on log loss
• D better than B on log loss
• F better than B on log loss
• F better than E on accuracy
• F better than E on log loss
• F better than C on accuracy
• F better than C on log loss

Held-out PubMedQA (never trained on):

• C better than A on ECE
• C better than A on log loss
• D better than A on ECE
• E better than A on log loss
• E better than B on log loss
• F better than A on ECE
• F better than A on log loss
• F better than B on log loss

In short

• Recommended target: F, 0.5 * one_hot(gold) + 0.5 * corrected_soft_label, with 2 dials, and an ensemble of 3 seeds if you can afford the training.
• Why F: against every target it was compared with (A, B, C, D and E), F was never reliably worse, on either split, under either check. On held-out it was reliably better than both gold-only targets on log loss. On the test split, it had the smallest seed-to-seed spread in accuracy and ECE.
• F against its closest rivals (D and E) is only partly settled. The average gap favoured F against D, in 5 of 6; against E, in 4 of 6 comparisons. Under both checks, F was better on: E test accuracy and E test log loss; the rest are within training luck.
• On data like the six training sources, one-hot gold with a dial is about as good as any soft target. soft_label alone (C) is the weakest there, because it inherits the teacher's mistakes.
• Transfer to the held-out domain is weak for every target. F's ensemble had the best held-out log loss (0.899, against 0.937 for predicting the label frequencies) and 0.595 accuracy (against 0.552 for always answering "yes").
• Judge targets on several seeds. Retraining alone moved one run's held-out accuracy by 0.076 (run A).

Teacher calibration on the test split

source	type	n	acc raw	acc calibrated	T	shift	fit	ECE raw	ECE calibrated
allenai/scicite	choice	214	0.799	0.799	5.46	+0.00	own	0.190	0.030
allenai/scicite	noul	86	0.674	0.744	4.89	+0.70	own	0.237	0.067
clinc/clinc_oos	choice	202	0.896	0.896	3.92	+0.00	own	0.096	0.067
clinc/clinc_oos	noul	98	0.776	0.867	3.22	+2.00	own	0.211	0.038
google-research-datasets/go_emotions	choice	216	0.569	0.569	8.86	+0.00	own	0.398	0.088
google-research-datasets/go_emotions	noul	84	0.619	0.738	9.46	+0.60	own	0.353	0.110
mteb/amazon_massive_intent	choice	212	0.901	0.901	5.46	+0.00	own	0.089	0.028
mteb/amazon_massive_intent	noul	85	0.859	0.871	4.10	+1.35	own	0.142	0.092
mteb/amazon_polarity	choice	190	0.963	0.963	5.46	+0.00	own	0.037	0.020
mteb/amazon_polarity	noul	110	0.955	0.955	1.98	-2.10	own	0.052	0.022
mteb/banking77	choice	216	0.921	0.921	5.11	+0.00	own	0.070	0.020
mteb/banking77	noul	84	0.905	0.905	4.68	-0.05	own	0.093	0.067

Sources

• CLINC150 (Larson et al., 2019), CC-BY-3.0
• GoEmotions (Demszky et al., 2020), Apache-2.0; soft labels from its raw per-rater annotations
• 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
• SciCite (Cohan et al., 2019), Apache-2.0 (authors' repository); downloaded from the authors' official link
• PubMedQA (Jin et al., 2019), expert-labelled PQA-L subset, MIT. Held out: never used for training or calibration.

Limitations

• Teacher soft labels reflect one 4-bit 7B teacher's judgement, calibrated on average per group.
• GoEmotions rater soft labels come from 3 raters per comment (5 when they disagreed), so they are coarse.
• Teacher-written option descriptions were reviewed by sampling, not one by one.
• Distractor difficulty depends on the embedding model (BAAI/bge-small-en-v1.5).
• English only. Gold labels inherit any noise in the source datasets.