README.md

metadata

configs:
  - config_name: math500
    data_files:
      - split: train
        path: data/math500.jsonl

math500

100 instances, evaluated from connections-dev/create-interpret-task-general-x100 (commit 8485bb7), eval code at git 87dd0ea. Files: data/math500.jsonl (one row per instance), metrics/math500.json, audit/math500.jsonl (answer groups that string match called wrong, next to the gold answer, to see what it misses).

Results

Metric	Value
Instances	100
Samples (successful / total)	9950 / 10000
Mean accuracy (pass@1)	0.959
pass@5	0.985
pass@10	0.988
pass@100	0.989
Majority vote	0.970
Truncated (hit the token limit)	0.0%
No answer	0.0%
Answer taken from \boxed{} (no tags)	99.7%
Answered without <answer> tags or a box	0.0%
Answer taken from the first bold span	0.0%
Confidently wrong (top wrong answer >= 50.0% of samples)	2.0%
Modal answer correct	97.0%
Neither (modal wrong, no dominant wrong answer)	1.0%
Distinct answers per instance (mean)	1.33
Answer entropy, bits (mean)	0.10

By level

level	Instances	Mean accuracy	Any sample correct
1	11	0.951	1.000
2	25	0.960	0.960
3	19	0.999	1.000
4	22	0.994	1.000
5	23	0.896	1.000

By subject

subject	Instances	Mean accuracy	Any sample correct
Algebra	25	1.000	1.000
Counting & Probability	6	0.973	1.000
Geometry	7	0.923	1.000
Intermediate Algebra	19	0.863	0.947
Number Theory	11	1.000	1.000
Prealgebra	18	0.958	1.000
Precalculus	14	0.999	1.000

How answers are scored

• Answer: the text after the last </think>. The last <answer>...</answer> wins (answer_source = tag). If the tags are missing, the last \boxed{...} is used (boxed; Qwen3's habit on math). Failing that, if the model wrote just the answer (one line, at most 3 tokens, generation finished normally), that is used and marked bare; failing that, the first **bold** span (at most 12 tokens) is used and marked bold (answers written as a sentence with the answer in bold). Anything else is no answer, which counts as wrong.
• Bucket: answers that mean the same thing are grouped: lowercase, strip accents/articles/punctuation, U.S. = United States, numbers (50,000, 50k, 8.3 million) and dates are canonicalised (a currency stays part of the number), and for SimpleQA an answer whose words are all inside exactly one longer answer is merged into it ("Coast Guard" into "United States Coast Guard"). Answers over 12 tokens are never merged. RAT is not merged.
• Match (correct): the bucket equals a gold answer (exact), or for SimpleQA is contained in it or contains it without hedging words (contains). Failed requests are null; no answer is false.
• Metrics: acc is the fraction of successful samples that are correct (= pass@1); pass@k is the unbiased estimator; majority vote takes the top bucket and splits ties as fractional credit; entropy is in bits over the answered buckets. Failed requests are excluded from every rate.
• Judge (SimpleQA only, optional): the simple-evals SimpleQA grader on each distinct bucket; per-sample verdicts are in the judge column.

Columns

Column	Description
instance_id	instance id (same ids as the generations repo)
gold	list of accepted gold answers
meta	dataset metadata for the instance (RAT: difficulty, cues; SimpleQA: topic, answer_type, ...)
n_samples / n_success / n_answered / n_bare / n_boxed	samples seen / requests that succeeded / samples with an answer / bare answers without tags / answers from \boxed{}
n_bold	answers taken from the first bold span of an untagged sentence
n_correct / n_truncated / n_no_answer	string-match correct / finish_reason=length / no answer
answer_counts	the distribution: list of {bucket, label, count, match, judge, members}, most common first
n_distinct / modal_answer / modal_frac / entropy	distinct buckets, top bucket, its share of successes, entropy in bits
modal_is_correct / top_wrong_frac / collapsed_wrong	top bucket correct; share of samples in the most common wrong bucket; that share >= 0.5 (confidently wrong)
acc / pass_any / majority_correct	fraction correct, any correct, majority vote (ties split)
judge_acc / judge_pass_any / judge_majority_correct / judge_counts	same, from the LLM judge (judge runs only)
sample_idx, ok, answer, answer_source, bucket, truncated, correct, judge	per-sample lists aligned by position: the judgment for every sample

Full metrics JSON

{
  "n_instances": 100,
  "n_instances_with_samples": 100,
  "n_samples": 10000,
  "n_success": 9950,
  "n_failed": 50,
  "acc_mean": 0.9592382820638425,
  "pass_any_mean": 0.99,
  "majority_correct_mean": 0.97,
  "truncation_rate": 0.00030150753768844223,
  "no_answer_rate": 0.00030150753768844223,
  "bare_rate": 0.0,
  "boxed_rate": 0.9968844221105527,
  "bold_rate": 0.0,
  "entropy_mean": 0.09660363697596018,
  "n_distinct_mean": 1.33,
  "collapse": {
    "threshold": 0.5,
    "collapsed_wrong_frac": 0.02,
    "modal_correct_frac": 0.97,
    "other_frac": 0.01,
    "top_wrong_frac_mean": 0.031058113226814156
  },
  "pass_at_k": {
    "1": {
      "value": 0.9592382820638425,
      "n_instances": 100
    },
    "5": {
      "value": 0.9846545271250211,
      "n_instances": 100
    },
    "10": {
      "value": 0.9883641964917858,
      "n_instances": 100
    },
    "100": {
      "value": 0.9893617021276596,
      "n_instances": 94
    }
  },
  "breakdowns": {
    "level": {
      "1": {
        "n_instances": 11,
        "acc_mean": 0.950909090909091,
        "pass_any_mean": 1.0
      },
      "2": {
        "n_instances": 25,
        "acc_mean": 0.96,
        "pass_any_mean": 0.96
      },
      "3": {
        "n_instances": 19,
        "acc_mean": 0.9994736842105262,
        "pass_any_mean": 1.0
      },
      "4": {
        "n_instances": 22,
        "acc_mean": 0.9936363636363637,
        "pass_any_mean": 1.0
      },
      "5": {
        "n_instances": 23,
        "acc_mean": 0.8962534002775766,
        "pass_any_mean": 1.0
      }
    },
    "subject": {
      "Algebra": {
        "n_instances": 25,
        "acc_mean": 0.9995999999999999,
        "pass_any_mean": 1.0
      },
      "Counting & Probability": {
        "n_instances": 6,
        "acc_mean": 0.9731786941580757,
        "pass_any_mean": 1.0
      },
      "Geometry": {
        "n_instances": 7,
        "acc_mean": 0.9228571428571428,
        "pass_any_mean": 1.0
      },
      "Intermediate Algebra": {
        "n_instances": 19,
        "acc_mean": 0.8629186166239164,
        "pass_any_mean": 0.9473684210526315
      },
      "Number Theory": {
        "n_instances": 11,
        "acc_mean": 1.0,
        "pass_any_mean": 1.0
      },
      "Prealgebra": {
        "n_instances": 18,
        "acc_mean": 0.958294573643411,
        "pass_any_mean": 1.0
      },
      "Precalculus": {
        "n_instances": 14,
        "acc_mean": 0.9992857142857143,
        "pass_any_mean": 1.0
      }
    }
  },
  "subset": "math500",
  "n_requested": "all",
  "input": {
    "repo": "connections-dev/create-interpret-task-general-x100",
    "sha": "8485bb7b97b7b1fa1474c31ea397b8aee0e9f5f7",
    "subset": "math500"
  },
  "enable_thinking": true,
  "answer_prefill": "",
  "judge_model": null,
  "judge_settings": null,
  "eval_git": {
    "sha": "87dd0ea592e6ef65a8bb3e5fa11fa92985654613",
    "dirty": false
  }
}