kbora/SoftElo

Soft-Elo Judge Annotations

Judge scores and preferences for From Uncertain Judgments to Calibrated Rankings: Conformal Elo Estimation for LLM Evaluation, Bora Kargi and David Salinas, NeurIPS 2026.

Code repository: kargibora/SoftElo. Dataset: kbora/SoftElo.

The release contains 368,777 judge–battle annotation rows from two public human-preference corpora. It supports alternative judge calibration, rating aggregation, committee and position-bias analyses without rerunning the judges. Each row corresponds to one judge's annotation of one battle. Counts below include invalid or incomplete scores; validity flags make the paper's filtering explicit.

Corpus	Judges	Annotation rows	Battles valid for every released judge
lmarena100k	8	199,922	18,158
lmarena140k	7	168,855	18,927

Annotation rows count repeated judgments of the same battle by different judges. The last column counts only the intersection across every released judge; it is not the number of battles available to each judge. All per-judge rows are released.

The annotations split is a storage name, not a training/test partition. Evaluation holds out models. Do not randomly split annotation rows and describe that as the paper's model-holdout evaluation.

Sources and attribution

• LMArena 100K: human preferences, model identities and source battle IDs.
• LMArena 140K: later Arena preferences and model identities.

The added annotations are LLM-generated scores and preferences, not additional human votes. Cite the upstream datasets and the Soft-Elo paper when using them. Source terms are described in DATA_LICENSES.md.

Loading

Load either LMArena configuration directly from the Hub. The dataset is public; no authentication is required.

from datasets import load_dataset

data = load_dataset(
    "kbora/SoftElo", name="lmarena100k", split="annotations", token=True
)

Use name="lmarena140k" for the second corpus. For reproducible analyses, pass revision= with the full dataset commit hash.

Use with the GitHub repository

The SoftElo repository accepts one judge's annotations as a CSV. From the loaded dataset, preserve the original row order and select the paper-valid rows:

frame = data.to_pandas()
frame = frame.loc[
    (frame["judge"] == "GPT-OSS-120B") & frame["paper_valid"]
].sort_values("row_index")
frame.to_csv("softelo_gpt_oss_120b.csv", index=False)

From a checkout of that repository:

python run_demo.py --data /path/to/softelo_gpt_oss_120b.csv --out outputs_gpt120

The CSV retains the score dictionaries as JSON strings, as expected by the GitHub loader. This example connects the data formats; the bundled reproduction package below documents the protocols for matching the archived paper tables.

The Python package loads a selected judge or paper subset, parses score dictionaries and applies the exact documented preference processing. It also provides commands to rebuild archived tables or recompute ratings.

Fields and orientation

Field	Meaning
battle_id	SHA-256 of corpus, source ID and sorted model pair; shared across judges and A/B orientation
source_id, source_revision	Join to the pinned upstream source record
source_dataset, judge	Corpus and paper judge identifier
instruction_id, row_index	Original experiment identity and row order
model_a, model_b	Model identities in this annotation's orientation
source_order_reversed	Whether annotation A/B is reversed relative to the linked source record
human_pref, judge_pref	Raw preferences: 0 = A wins, 1 = B wins, 0.5 = tie; weak values 0.25/0.75 are retained
scores_a, scores_b	Aggregated six-criterion scores, encoded as JSON objects
scores_a_original, scores_b_original	Scores with A presented first, where retained
scores_a_swapped, scores_b_swapped	Scores with B presented first, mapped to the same model A/B identities
preference_original, preference_swapped	Model-oriented preferences for each presentation order, where retained
has_order_scores, has_order_preferences	Availability of those per-order fields
paper_valid	Original per-judge score-validity rule: dictionaries with no missing/nonfinite score values
shared_eligible	Stricter shared-support rule: finite scores in [1,10], valid preference labels and no parse error
language, len_a, len_b, parse_error	Available annotation metadata

Criteria are adherence, clarity, completeness, factuality, fluency and helpfulness. JSON objects preserve missing score values as null. Unavailable per-order objects are null, not fabricated duplicates of the aggregate. Five older LMArena100K judge files retain only aggregated scores; see per-judge availability counts in manifest.json. Metadata files contain the saved system prompt, rubric and available judge settings; unknown settings remain null.

Paper subsets

paper_valid is the filter used for the per-judge main rating analysis. The reference loader subsequently maps preferences within 0.26 of 0.5 to ties.

subsets/*_shared_valid.parquet lists the intersection of eligible source records and model pairs across all released judges in each corpus. Human preferences agree after aligning A/B orientation. The LMArena100K intersection contains exactly the 18,158 battles used by the paper's committee and human reference analyses. The LMArena140K intersection is supplied for reuse; it is not a claim that the paper evaluated all judges on that intersection.

Preserve all per-judge rows for reproducing the main results. Selecting only the shared intersection changes the experiment.

Source versions

Every exported source ID, model pair and human label was checked against a pinned upstream snapshot. Use each row's source_revision when joining.

The original annotation files did not record upstream commit IDs. These are verified compatible snapshots, not a reconstruction of unrecorded run metadata. Battle identifiers match records across versions by source ID and model pair; exact historical conversation-text equality across versions has not been independently checked. Original text can be fetched from the pinned upstream sources with load_source_rows in the Python package.

Reproduction and limits

The package provides explicit protocols for the main fold-specific-temperature analysis, pooled-temperature ablations, constant smoothing and the historical conformal calculation. Archived main model-level predictions and conformal split measurements are included under results/. The smoothing archive contains per-judge MAEs; its original model-level predictions were not saved.

Fresh CPU fits reproduce all 31 judge/protocol combinations: eight main, eight smoothing, eight historical conformal and seven LMArena140K comparisons. See code/VERIFICATION.md for the reference scope and machine-readable checks. The audit corrected Hard-Elo anchor bootstrapping in the release code; the saved reference results are unchanged. This release does not include auxiliary prompt-variation experiments or the author re-annotation study.

The scores reflect judge errors and biases. Human votes are observed preferences, not noise-free ground truth. Calibration and conformal coverage may change under model-population or prompt-distribution shifts. Missing per-order fields limit position-bias analyses for those judges. This is an annotation dataset; it does not include raw user prompts, candidate responses, user identifiers or judge explanations.

Citation

@inproceedings{kargi2026softelo,
  title = {From Uncertain Judgments to Calibrated Rankings: Conformal Elo Estimation for LLM Evaluation},
  author = {Kargi, Bora and Salinas, David},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2026}
}

For Arena, also cite Chiang et al., Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference (2024), and Tang, Chiang and Angelopoulos, Arena Explorer: A Topic Modeling Pipeline for LLM Evals & Analytics (2025), as requested by the source dataset card.