LIQIIIII/Jeva-Eval

Jeva Evaluation Suite

Prepared evaluation inputs for Jeva: System One Models from First Principles.

Paper · Code · Model · Project

This private repository contains the eight general-evaluation task slices, the original and nested ToolRet candidate-count diagnostics, and the MLDR long-document inputs and cache workload. Candidate text, ordering, and targets are frozen to the evaluated inputs. It contains evaluation data only, not model weights or training recipes.

Contents

Configuration	Records	Split
typed_test	2,000	test
banking77_test_K8	3,080	test
banking77_test_K16	3,080	test
massive_test_K8	2,974	test
massive_test_K16	2,974	test
mind2web_cross_task_K16_present	1,974	test
mind2web_cross_website_K16_present	1,314	test
mind2web_cross_domain_K16_present	5,590	test
toolret_K32	84	evaluation
toolret_K128	84	evaluation
toolret_K256	83	evaluation
toolret_K512	83	evaluation
toolret_K1024	83	evaluation
toolret_K2048	83	evaluation
toolret_nested_K32	83	evaluation
toolret_nested_K128	83	evaluation
toolret_nested_K256	83	evaluation
toolret_nested_K512	83	evaluation
toolret_nested_K1024	83	evaluation
toolret_nested_K2048	83	evaluation
mldr	200	evaluation
mldr_cache50	50	evaluation

The eight general tasks contain 22,986 decisions. Typed Decisions includes 600 Choice, 600 Noul, and 800 Score decisions. BANKING77 and MASSIVE each have 8- and 16-candidate slices. Mind2Web has cross-task, cross-website, and cross-domain slices with 16 supplied candidates; these measure candidate selection.

The original ToolRet diagnostic contains 500 distinct queries distributed across six candidate counts. The nested version evaluates the same 83 queries at all six counts, producing 498 records. It retains supplied positives and uses weak or unjudged distractors. These are prepared ToolRet-derived diagnostics, not official ToolRet test leaderboard submissions.

MLDR contains 200 English long-document requests. The mldr_cache50 configuration is the exact 50-query subset used for the cache measurements. protocols/mldr_cache50.json contains its 200 episodes: one shared-bank condition and three novel-set overlap conditions per query. Prewarming uses different query states; its cost is measured separately. Load mldr, not only mldr_cache50, when restoring priming inputs: some priming states and texts come from the larger retained pool. MLDR here is a diagnostic derived from upstream training-split data, not the official MLDR test set.

Download

Access requires a Hugging Face account authorized to read this private repository.

hf auth login
hf download LIQIIIII/Jeva-Eval --repo-type dataset --local-dir ./jeva-eval

Every task is a self-contained gzip-compressed JSONL file with inline text; no external text database is needed. manifest.json lists counts, candidate counts, scoring rules, and file hashes. SHA256SUMS verifies the downloaded files:

cd jeva-eval
sha256sum -c SHA256SUMS

For reproducible downloads, supply --revision with the dataset commit hash.

Load a task

from datasets import load_dataset

typed = load_dataset("LIQIIIII/Jeva-Eval", "typed_test", split="test", token=True)
large = load_dataset("LIQIIIII/Jeva-Eval", "toolret_nested_K2048", split="evaluation", token=True)

Use an immutable revision for comparisons. The split name evaluation denotes the prepared diagnostic, not an upstream official test split.

Evaluate locally

After installing the Jeva package, pass a downloaded task directly to the evaluator:

jeva-evaluate \
  --model ./models/Jeva-4B \
  --data ./jeva-eval/data/typed_test.jsonl.gz \
  --output ./results/typed_test

Here ./models/Jeva-4B is a downloaded local checkpoint. For a Hub model ID, also pass --revision with its immutable 40-character commit hash. Set the state and option token limits explicitly for long texts and reproduce the paper's batching and cache settings when comparing latency. Downloading these inputs does not select a model checkpoint or an execution protocol automatically.

To restore the MLDR cache workload using the supplied texts:

import json
from pathlib import Path
from jeva.evaluation import read_jsonl, hydrate_cache_episodes

root = Path("./jeva-eval")
plan = json.loads((root / "protocols/mldr_cache50.json").read_text())
records = read_jsonl(root / "data/mldr.jsonl.gz")
episodes = hydrate_cache_episodes(plan["episodes"], records)

These episodes can be supplied to evaluate_cache_episodes in the Jeva package. Reset caches between episodes and modes. No previous answer is reused.

Record format and scoring

• id, group_id: frozen decision and optional grouping identifiers.
• state, question, decision_type, options: model inputs; option order is significant for exact reproduction.
• allow_null: whether candidate absence is a permitted outcome.
• target_probs: null probability first, followed by the probability for each option. Soft targets are retained.
• positive_index: zero-based positive Boolean candidate for Noul; null for other types.
• values: numerical rubric values for Score; null for other types.
• option_ids: candidate identifiers aligned with the option texts. Generated text hashes are used where the original format had no identifiers.
• accuracy_rule: first_argmax for the general suite; tied_top_set for the supplied ToolRet and MLDR diagnostics.
• source: the configuration containing the record.

The general suite uses first-argmax correctness, including null when applicable. Diagnostic tied-top correctness accepts a predicted candidate whose target probability is within 1e-8 of the maximum. Noul uses soft positive-candidate targets; Score uses expected rubric values in their original units. Preserve the target distributions when computing fidelity metrics.

Candidate-order evaluation uses the 600 Choice rows in typed_test, with original, reversed, and up to three distinct seeded random orders. The frozen settings are in protocols/typed_permutations.json; align distributions by candidate identity. Nested ToolRet queries repeat across K, and the MLDR cache subset overlaps mldr; do not sum configuration sizes as unique examples.

Sources and terms

Source attribution, links, licenses, and transformation notes are recorded in LICENSES.md. Each component retains its upstream terms; no single blanket license is applied to third-party text. This is a private research archive. ToolRet redistribution terms and the underlying long-document text provenance require review before any public release.

Citation

@misc{li2026jeva,
  title={Jeva: System One Models from First Principles},
  author={Qi Li and Xinchao Wang},
  year={2026},
  eprint={2610.18773},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2610.18773}
}

Please also credit the original datasets when using their corresponding configurations.