Datasets:

cs-giung
/

base-evals

language:
  - en
license: other
task_categories:
  - multiple-choice
tags:
  - evaluation
  - base-models
  - zero-shot
  - log-likelihood
configs:
  - config_name: default
    data_files:
      - split: tiny_arithmark_2_0
        path: data/tiny_arithmark_2_0-00000-of-00001.parquet
      - split: tiny_arithmark_3_0
        path: data/tiny_arithmark_3_0-00000-of-00001.parquet
      - split: tiny_hellaswag
        path: data/tiny_hellaswag-00000-of-00001.parquet
      - split: tiny_logicmark
        path: data/tiny_logicmark-00000-of-00001.parquet
      - split: tiny_openbookqa
        path: data/tiny_openbookqa-00000-of-00001.parquet
      - split: tiny_pythonio
        path: data/tiny_pythonio-00000-of-00001.parquet
      - split: tiny_truthfulqa
        path: data/tiny_truthfulqa-00000-of-00001.parquet

base-evals

0-shot, 4-choice continuation evaluation for pretrained base models.

Splits

split	source	source split	rows
`tiny_arithmark_2_0`	`AxiomicLabs/ArithMark-2.0@5bd986c7`	`train`	500
`tiny_arithmark_3_0`	`AxiomicLabs/Arithmark-3.0@6f6e59dd`	`train`	500
`tiny_hellaswag`	`Rowan/hellaswag@218ec52e`	`validation`	500
`tiny_logicmark`	`AxiomicLabs/LogicMark@e357c145`	`train`	500
`tiny_openbookqa`	`allenai/openbookqa@388097ea`	`test`	500
`tiny_pythonio`	`XiangPan/pythonio-mc@a85256a1`	`test`	500
`tiny_truthfulqa`	`EleutherAI/truthful_qa_mc@b6468d07`	`validation`	500
total			3,500

Format

id: string, context: string, choices: list[string], label: int64. Labels are zero-based. Input is context + choices[i], with exactly one ASCII space before each choice and no trailing context whitespace. PythonIO uses LF line endings, Input: {input}\nOutput: contexts, and choices without outer trailing whitespace; code indentation and output-value whitespace are retained. HellaSwag uses context cleanup; TruthfulQA uses Q: {question}\nA:.

Evaluation

No few-shot examples, chat template, CoT, or added BOS/EOS.

acc: choose the largest sum of continuation-token log-probabilities.
acc_norm: choose the largest mean log-probability per continuation token.

Exclude context/padding; include boundary whitespace. Reject ambiguous token boundaries and overlength inputs. Break ties by first index. Report each split separately. This custom protocol is not HellaSwag character normalization, TruthfulQA MC2, or IRT.

Provenance

Each split selects 500 questions with random.Random(42); OpenBookQA retains its full test set. ArithMark 3.0 excludes time_session prompts without duration information before sampling. Rows with fewer than four choices are excluded; larger sets keep the answer and three sampled incorrect choices. Source row/choice order is preserved. Full rules, hashes and indices: manifest.json; original inputs/cards: sources/; offline rebuild: build.py.

Evaluation only, including upstream train splits. Original source terms apply; PythonIO and OpenBookQA have no declared license. Pretraining contamination is unknown.