Dataset cartography

Overview

Dataset cartography scores each training example by its training dynamics: how the model's probability for an example changes from one saved epoch to the next. Examples whose score is high are treated as ambiguous. Examples whose score is low are treated as easy. The WANLI paper uses the "variability" version of this idea, meaning the spread of the probability of the correct label across epochs, to pick which examples to build on 2201.05955 §2.1.

Only one card on this wiki uses the technique, so every claim below rests on WANLI. The original cartography paper is named in WANLI's lineage but is not compiled here 2201.05955 Lineage, so its own results are not on this page.

Numbers are author-reported, single-run unless a card says otherwise.

Variants

Gold-label variability

For each training example, record the probability the model gives its gold label after each saved epoch. The variability is the standard deviation of those probabilities. The WANLI paper calls an example ambiguous when this variability is high 2201.05955 §2.1. The authors take the most ambiguous 25% of each gold label as seed examples for generation, and the telephone genre is excluded from that step 2201.05955 §2.1. Card: WANLI.

Max variability for unlabelled candidates

Generated candidates have no gold label that the model was trained on, so WANLI scores them with an estimator. For each candidate, the score is the largest, across the three labels, of the standard deviation of that label's probability over the saved epochs 2201.05955 eq. 1. The authors state that this assumes adding the candidate to training would barely change any saved checkpoint 2201.05955 §2.3. Card: WANLI.

Per-label quota selection

Rank by score inside each label, then keep a fixed share from each label. The WANLI filter keeps half of the filtered set and gives each label the same count, so the classes stay balanced 2201.05955 §2.3. Balancing is only approximate in the final data: contradictions are about 15% of the labelled examples 2201.05955 §5.3.

When to use which

Implementation sketch

Pseudocode in our own words, with the paper section for each step. Assumes a list of per-epoch probabilities for each example.

# probs[e][y]: model probability of label y after saved epoch e.
# Requires one saved checkpoint per epoch [arxiv:2201.05955 §2.3].

def gold_variability(probs, gold):
    # Std of the gold-label probability across epochs; high = ambiguous
    # [arxiv:2201.05955 §2.1]
    return std([probs[e][gold] for e in range(num_epochs)])

def max_variability(probs):
    # No gold label: largest per-label std across epochs [arxiv:2201.05955 eq. 1]
    return max(std([probs[e][y] for e in range(num_epochs)])
               for y in LABELS)

def select_per_label(examples, score, keep_fraction):
    # Keep the highest-scoring share inside each label, so classes stay
    # balanced [arxiv:2201.05955 §2.3]
    chosen = []
    for y in LABELS:
        pool = sorted(of_label(examples, y), key=score, reverse=True)
        chosen += pool[: int(len(pool) * keep_fraction)]
    return chosen

Use population or sample standard deviation consistently, because the paper does not say which 2201.05955 §2.3; see Pitfalls.

Contradictions

The two WANLI-alone rows in the same paper do not match. Table 3 reports QNLI 81.40, HANS 89.28, BIG-Bench 85.19 and WNLI 67.28 2201.05955 tab. 3. Table 10 reports the same training set, at the same size, with QNLI 76.81, HANS 89.40, BIG-Bench 75.40 and WNLI 65.15 2201.05955 tab. 10. The paper does not explain the gap.

Pitfalls

• Model-dependent scores. Seeds, neighbours and scores all depend on one specific model M, so changing M changes the selected data 2201.05955 §2.1.
• Checkpoints are required. Without one saved checkpoint per epoch the estimator cannot be computed as the paper describes 2201.05955 §2.3.
• Weak estimator. The max-variability estimator correlates only 0.527 with true variability on one held-out sample 2201.05955 app. A.
• Standard deviation convention. The paper names "std" without a denominator 2201.05955 §2.3. The card on this wiki notes that the released code uses the population form, so compare with the same convention.
• Single runs. Each setting is one training run 2201.05955 app. B. Differences of about one or two points need a variance estimate before they count as real.
• Label imbalance after review. A balanced selection did not give balanced labels after human review 2201.05955 §5.3. Plan for class weighting or rebalancing when comparing against other sets.
• Label noise. On the disagreement cases the paper examined, two authors agreed on the label only 49% of the time 2201.05955 app. D.4.
• Generation access. WANLI's candidate generation used a GPT-3 model that was available through an API in 2021. Reproducing that step depends on comparable model access 2201.05955 app. C.2.

Papers

• WANLI: Worker and AI Collaboration for Natural Language Inference Dataset Creation: uses training-dynamics variability to pick seed examples and to filter generated NLI examples, and reports the resulting dataset's out-of-domain accuracy.