File size: 3,622 Bytes
0e25b54 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 | ---
pretty_name: Curriculum Contamination Coherent-Wrong-Label Tasks (Stage 1)
language:
- en
task_categories:
- question-answering
- text-generation
tags:
- curriculum-learning
- self-training
- mathematical-reasoning
- synthetic-data
size_categories:
- n<1K
configs:
- config_name: unique_tasks
data_files:
- split: train
path: data/unique_tasks.jsonl
- config_name: label_twins
data_files:
- split: train
path: data/label_twins.jsonl
---
# Curriculum Contamination Coherent-Wrong-Label Tasks (Stage 1)
This private pre-release contains a controlled mathematical task bank for
studying whether self-evolving curriculum selectors can admit an objectively
wrong answer when repeated solver samples form a coherent majority.
## Dataset contents
- `data/unique_tasks.jsonl`: 240 unique, exact-answer mathematical tasks.
- `data/label_twins.jsonl`: 480 surface-identical task/reference pairs: one
objectively correct reference and one plausible but objectively incorrect
reference for each task.
- `metadata/validation_report.json`: construction and validation summary.
- `metadata/*.integrity.json`: line-count and SHA-256 integrity records.
The 240 tasks cover ten families, with 24 tasks per family:
1. area scaling;
2. at-least-one probability;
3. Bayesian positive-test reasoning;
4. combined work rates;
5. equal-distance average speed;
6. recurrence indexing;
7. repeated-object permutations;
8. successive percentage change;
9. weighted means;
10. probability without replacement.
## Important interpretation boundary
The plausible wrong answer encodes a prespecified candidate shortcut. It is not
claimed to be a model error until an evaluator run independently observes it.
Fields such as `candidate_shortcut_not_observed_model_behavior` and
`selector_observation_status: not_yet_run` preserve this distinction.
This Stage 1 release contains no completed selector outcome, natural-prevalence
estimate, training-effect estimate, or closed-loop contamination result.
## Construction and validation
- All 240 tasks have a unique exact answer under an independent deterministic
verifier.
- The plausible wrong answer differs from the objective answer for every task.
- All 240 correct/wrong pairs have identical question text.
- Thirty tasks were included in a blinded machine audit during construction.
- The persisted validation report records 240/240 independent oracle passes and
no construction errors.
## Schema overview
`unique_tasks.jsonl` includes the question, family, parameters, objective answer,
plausible wrong answer, misconception identifier and verifier certificate.
`label_twins.jsonl` includes the shared question and one proposed reference per
row, with `condition` and `reference_correctness` distinguishing the pair.
## Intended uses
- diagnostic evaluation of majority-vote pseudo-labeling;
- controlled curriculum-admission studies;
- comparison of selector risk/retention frontiers;
- paired causal studies of correct versus coherent-wrong training updates.
## Limitations
- The pool is controlled and deliberately enriched for candidate shortcuts; it
does not estimate their prevalence in naturally generated curricula.
- It currently covers exact-answer English mathematical tasks only.
- Candidate wrong answers are mechanism probes, not labels of observed model
behavior.
- Downstream learning harm must be established by a separate intervention.
## License and citation
No public-use license or paper citation is assigned in this private pre-release.
They will be added only after an explicit owner decision and double-blind release
review.
|