Hugging Face's logo

Datasets:
chaoliangUNSW
/
decision-index-results

Tags:
decision-index
jev-style
benchmark-results
License:
Dataset card Files
xet
 Community
Dataset Viewer
Duplicate
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
parts: struct<jev-style-0.8b-decision-v3: struct<engine: string, engine_options: struct<model: string, revi (... 1467 chars omitted)
  child 0, jev-style-0.8b-decision-v3: struct<engine: string, engine_options: struct<model: string, revision: string>, model_source: struct (... 654 chars omitted)
      child 0, engine: string
      child 1, engine_options: struct<model: string, revision: string>
          child 0, model: string
          child 1, revision: string
      child 2, model_source: struct<model: string, hf_repo: string, revision: string, snapshot: string, runtime_sha256: string, m (... 282 chars omitted)
          child 0, model: string
          child 1, hf_repo: string
          child 2, revision: string
          child 3, snapshot: string
          child 4, runtime_sha256: string
          child 5, manifest_verified: struct<checked: int64, bad: list<item: null>, missing_ignored: list<item: null>>
              child 0, checked: int64
              child 1, bad: list<item: null>
                  child 0, item: null
              child 2, missing_ignored: list<item: null>
                  child 0, item: null
          child 6, device: string
          child 7, dtype: string
          child 8, template: string
          child 9, context_limit: int64
          child 10, head_max: int64
          child 11, temperature: string
          child 12, no_truncation: bool
          child 13, option_filtering: bool
          child 14, per_benchmark_prompts: bool
      child 
...
 struct<metric: string, score: double, scored_requests: int64>
              child 0, metric: string
              child 1, score: double
              child 2, scored_requests: int64
          child 3, GSM8K-4choice: struct<metric: string, score: double, scored_requests: int64>
              child 0, metric: string
              child 1, score: double
              child 2, scored_requests: int64
          child 4, iSarcasmEval-A-Ar: struct<metric: string, score: null, scored_requests: int64>
              child 0, metric: string
              child 1, score: null
              child 2, scored_requests: int64
          child 5, iSarcasmEval-A-En: struct<metric: string, score: null, scored_requests: int64>
              child 0, metric: string
              child 1, score: null
              child 2, scored_requests: int64
          child 6, iSarcasmEval-B-En: struct<metric: string, score: null, scored_requests: int64>
              child 0, metric: string
              child 1, score: null
              child 2, scored_requests: int64
          child 7, iSarcasmEval-C-Ar: struct<metric: string, score: null, scored_requests: int64>
              child 0, metric: string
              child 1, score: null
              child 2, scored_requests: int64
          child 8, iSarcasmEval-C-En: struct<metric: string, score: null, scored_requests: int64>
              child 0, metric: string
              child 1, score: null
              child 2, scored_requests: int64
engine: string
to
{'engine': Value('string'), 'counts': {'ok': Value('int64'), 'unsupported': Value('int64')}, 'successful_request_latency_ms': {'median': Value('float64'), 'p95': Value('float64'), 'mean': Value('float64')}, 'benchmarks': List({'catalog_id': Value('int64'), 'dataset': Value('string'), 'requests': Value('int64'), 'answered': Value('int64'), 'unsupported': Value('int64'), 'errors': Value('int64'), 'abstained': Value('int64'), 'pending': Value('int64'), 'scored_requests': Value('int64'), 'metric': Value('string'), 'score': Value('float64'), 'reference_same_cases': Value('null'), 'median_ms': Value('float64'), 'detail': {'field_accuracy': Value('float64'), 'case_exact_accuracy': Value('float64'), 'custom_metrics': {'ndcg_at_10': Value('float64'), 'mrr': Value('float64'), 'recall_at_10': Value('float64'), 'candidate_recall': Value('float64'), 'scorable_candidate_recall': Value('float64'), 'candidates_scored': Value('int64'), 'candidates_retrieved': Value('int64'), 'bm25_ndcg_at_10': Value('float64'), 'quality_quality': Value('float64'), 'quality_cost_usd': Value('float64'), 'quality_utility': Value('float64'), 'quality_oracle_optimal': Value('float64'), 'quality_utility_regret': Value('float64'), 'cost_aware_quality': Value('float64'), 'cost_aware_cost_usd': Value('float64'), 'cost_aware_utility': Value('float64'), 'cost_aware_oracle_optimal': Value('float64'), 'cost_aware_utility_regret': Value('float64'), 'value_regret': Value('float64'), 'brier': Value('float64'), 'log_loss': Va
...
lue('int64'), 'mean': Value('float64')}}, 'cluster_macro_accuracy': Value('float64'), 'positive_f1_by_field': {'sarcastic': Value('float64'), 'sarcasm': Value('float64'), 'irony': Value('float64'), 'satire': Value('float64'), 'understatement': Value('float64'), 'overstatement': Value('float64'), 'rhetorical_question': Value('float64')}, 'category_macro_f1': Value('float64'), 'scored_fields': Value('int64'), 'chance_on_rows': Value('float64')}, 'tracks': {'RouterBench-0shot': {'metric': Value('string'), 'score': Value('float64'), 'scored_requests': Value('int64')}, 'RouterBench-5shot': {'metric': Value('string'), 'score': Value('float64'), 'scored_requests': Value('int64')}, 'GSM8K-10choice': {'metric': Value('string'), 'score': Value('float64'), 'scored_requests': Value('int64')}, 'GSM8K-4choice': {'metric': Value('string'), 'score': Value('float64'), 'scored_requests': Value('int64')}, 'iSarcasmEval-A-Ar': {'metric': Value('string'), 'score': Value('null'), 'scored_requests': Value('int64')}, 'iSarcasmEval-A-En': {'metric': Value('string'), 'score': Value('null'), 'scored_requests': Value('int64')}, 'iSarcasmEval-B-En': {'metric': Value('string'), 'score': Value('null'), 'scored_requests': Value('int64')}, 'iSarcasmEval-C-Ar': {'metric': Value('string'), 'score': Value('null'), 'scored_requests': Value('int64')}, 'iSarcasmEval-C-En': {'metric': Value('string'), 'score': Value('null'), 'scored_requests': Value('int64')}}}), 'note': Value('string'), 'edition': Value('string')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 544, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3052, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 558, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 361, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2381, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              parts: struct<jev-style-0.8b-decision-v3: struct<engine: string, engine_options: struct<model: string, revi (... 1467 chars omitted)
                child 0, jev-style-0.8b-decision-v3: struct<engine: string, engine_options: struct<model: string, revision: string>, model_source: struct (... 654 chars omitted)
                    child 0, engine: string
                    child 1, engine_options: struct<model: string, revision: string>
                        child 0, model: string
                        child 1, revision: string
                    child 2, model_source: struct<model: string, hf_repo: string, revision: string, snapshot: string, runtime_sha256: string, m (... 282 chars omitted)
                        child 0, model: string
                        child 1, hf_repo: string
                        child 2, revision: string
                        child 3, snapshot: string
                        child 4, runtime_sha256: string
                        child 5, manifest_verified: struct<checked: int64, bad: list<item: null>, missing_ignored: list<item: null>>
                            child 0, checked: int64
                            child 1, bad: list<item: null>
                                child 0, item: null
                            child 2, missing_ignored: list<item: null>
                                child 0, item: null
                        child 6, device: string
                        child 7, dtype: string
                        child 8, template: string
                        child 9, context_limit: int64
                        child 10, head_max: int64
                        child 11, temperature: string
                        child 12, no_truncation: bool
                        child 13, option_filtering: bool
                        child 14, per_benchmark_prompts: bool
                    child 
              ...
               struct<metric: string, score: double, scored_requests: int64>
                            child 0, metric: string
                            child 1, score: double
                            child 2, scored_requests: int64
                        child 3, GSM8K-4choice: struct<metric: string, score: double, scored_requests: int64>
                            child 0, metric: string
                            child 1, score: double
                            child 2, scored_requests: int64
                        child 4, iSarcasmEval-A-Ar: struct<metric: string, score: null, scored_requests: int64>
                            child 0, metric: string
                            child 1, score: null
                            child 2, scored_requests: int64
                        child 5, iSarcasmEval-A-En: struct<metric: string, score: null, scored_requests: int64>
                            child 0, metric: string
                            child 1, score: null
                            child 2, scored_requests: int64
                        child 6, iSarcasmEval-B-En: struct<metric: string, score: null, scored_requests: int64>
                            child 0, metric: string
                            child 1, score: null
                            child 2, scored_requests: int64
                        child 7, iSarcasmEval-C-Ar: struct<metric: string, score: null, scored_requests: int64>
                            child 0, metric: string
                            child 1, score: null
                            child 2, scored_requests: int64
                        child 8, iSarcasmEval-C-En: struct<metric: string, score: null, scored_requests: int64>
                            child 0, metric: string
                            child 1, score: null
                            child 2, scored_requests: int64
              engine: string
              to
              {'engine': Value('string'), 'counts': {'ok': Value('int64'), 'unsupported': Value('int64')}, 'successful_request_latency_ms': {'median': Value('float64'), 'p95': Value('float64'), 'mean': Value('float64')}, 'benchmarks': List({'catalog_id': Value('int64'), 'dataset': Value('string'), 'requests': Value('int64'), 'answered': Value('int64'), 'unsupported': Value('int64'), 'errors': Value('int64'), 'abstained': Value('int64'), 'pending': Value('int64'), 'scored_requests': Value('int64'), 'metric': Value('string'), 'score': Value('float64'), 'reference_same_cases': Value('null'), 'median_ms': Value('float64'), 'detail': {'field_accuracy': Value('float64'), 'case_exact_accuracy': Value('float64'), 'custom_metrics': {'ndcg_at_10': Value('float64'), 'mrr': Value('float64'), 'recall_at_10': Value('float64'), 'candidate_recall': Value('float64'), 'scorable_candidate_recall': Value('float64'), 'candidates_scored': Value('int64'), 'candidates_retrieved': Value('int64'), 'bm25_ndcg_at_10': Value('float64'), 'quality_quality': Value('float64'), 'quality_cost_usd': Value('float64'), 'quality_utility': Value('float64'), 'quality_oracle_optimal': Value('float64'), 'quality_utility_regret': Value('float64'), 'cost_aware_quality': Value('float64'), 'cost_aware_cost_usd': Value('float64'), 'cost_aware_utility': Value('float64'), 'cost_aware_oracle_optimal': Value('float64'), 'cost_aware_utility_regret': Value('float64'), 'value_regret': Value('float64'), 'brier': Value('float64'), 'log_loss': Va
              ...
              lue('int64'), 'mean': Value('float64')}}, 'cluster_macro_accuracy': Value('float64'), 'positive_f1_by_field': {'sarcastic': Value('float64'), 'sarcasm': Value('float64'), 'irony': Value('float64'), 'satire': Value('float64'), 'understatement': Value('float64'), 'overstatement': Value('float64'), 'rhetorical_question': Value('float64')}, 'category_macro_f1': Value('float64'), 'scored_fields': Value('int64'), 'chance_on_rows': Value('float64')}, 'tracks': {'RouterBench-0shot': {'metric': Value('string'), 'score': Value('float64'), 'scored_requests': Value('int64')}, 'RouterBench-5shot': {'metric': Value('string'), 'score': Value('float64'), 'scored_requests': Value('int64')}, 'GSM8K-10choice': {'metric': Value('string'), 'score': Value('float64'), 'scored_requests': Value('int64')}, 'GSM8K-4choice': {'metric': Value('string'), 'score': Value('float64'), 'scored_requests': Value('int64')}, 'iSarcasmEval-A-Ar': {'metric': Value('string'), 'score': Value('null'), 'scored_requests': Value('int64')}, 'iSarcasmEval-A-En': {'metric': Value('string'), 'score': Value('null'), 'scored_requests': Value('int64')}, 'iSarcasmEval-B-En': {'metric': Value('string'), 'score': Value('null'), 'scored_requests': Value('int64')}, 'iSarcasmEval-C-Ar': {'metric': Value('string'), 'score': Value('null'), 'scored_requests': Value('int64')}, 'iSarcasmEval-C-En': {'metric': Value('string'), 'score': Value('null'), 'scored_requests': Value('int64')}}}), 'note': Value('string'), 'edition': Value('string')}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Decision Index runs of Jev-Style models

Self-scored, complete runs of the Decision Index public suites, scored with the kit's own scorer. Every text run's scores.json and every vision run's vision-scores.json says "complete": true. results.jsonl.gz is the kit's compact format (no payload, image or raw output), so the files carry no benchmark text (GPQA and HLE ask that their questions not be redistributed).

On the public board. Since 2026-11-04 the Decision Index board (edition 0.3) lists jev-style-2b-v4 at Public 44.37, Full 34.53 (the Full score is from private tests run by the maintainers), Breadth 41.75, replacing jev-style-2b-v3; the Vision board lists it at Public 57.55 / Full 51.94. jev-style-0.8b-v3 is listed at Public 17.0 / Full 12.5. The Public numbers are the ones from the runs below.

Jev-Style 2B v4: Decision Index 0.3 public suite (text) and the public Vision suite

Run Public index Raw index Knowledge Language Retrieval Tools Arts Requests Scores
Jev-Style-2B-Decision-v4 @ 3f2c063f, text 44.37 57.68 23.6 42.5 65.4 61.8 29.0 140,178 scoreable, all answered (0 unsupported) scores.json
Run Public vision index Rows Scores
Jev-Style-2B-Decision-v4 @ 3f2c063f, vision 57.55 12,210, all answered vision-scores.json
  • Engines: macjev.eval.di_v4:V4TextEngine and macjev.eval.di_v4:V4VisionEngine, shipped in the model repo under src/ (the scorer the model was trained, calibrated and evaluated with). Kit 25c6ec7 (edition 0.3), its runners and scorers unmodified. bf16 autocast, fp32 readout, one calibrated temperature per question type (readout_config.json), no text generated, nothing truncated. Text: fast path on (each request's state computed once and cached, its questions batched). Vision: every image at native resolution up to 1.6 MP.
  • Hardware: one NVIDIA A100-SXM4-80GB (Google Colab). The text run is 5 parallel shards (sha256(group_id) mod 5); the vision run is the vision pipeline run plus the rows not yet answered split by sha256(row id) mod N; the processes shared the GPU through NVIDIA MPS, so the per-request times in the results include sharing. Every run_id appears exactly once; environment.json lists every part.
  • Latency without sharing (one process, one request at a time, 750 text requests sampled uniformly from the suite files, the model repo's code only, bf16): with CUDA graphs (--option cuda_graphs=true) median 16.4 ms, mean 58.8 ms, p80 60.9 ms, p95 211 ms; with the fast path alone (the setting of the full run) 71.5 / 104.0 / 80.5 / 214 ms. 663 of the 750 requests are in the run: with the fast path their answers matched the full run on all 1,585 of their questions, with CUDA graphs 1,579 of the 1,585 have the same top option.
  • Suites rebuilt from the pinned sources and verified. Text: suite rebuild on 25c6ec7 no longer reproduces the pinned release-v1 hash (PR #165 changed the SGD rows without updating the pins), so the SGD builder line from before #165 was patched in-process (kit files unmodified); all four pins then matched. Vision: rebuilt with Pillow 12.3.0 (the kit's documented version) and verified.
  • Model identity: the engine loaded the released weights. The run label v4-main-colab-seed1 in response.model and environment.json is the label written in the model repository's macjev_readout.json; weights_digest d03c41b67e4205b3 is the sha256 over the SHA256SUMS.json entries of the three weight shards, the same for revision 3f2c063f.
  • These runs used the main repository's PyTorch runtime. The text-only GGUF and MLX builds were not run on the index; their agreement with this runtime is on their cards.
  • Several benchmarks' official train splits are in v4's training data (listed on the model card, including CORD of the Vision KIE benchmark), so those results are not zero-shot.

Jev-Style v3 (0.8B and 2B): Decision Index 0.2.1

The 2B v3 was listed on the public board (Public index 34.05 on the 0.4.1 board) until 2026-11-04, when the 2B v4 replaced it; the 0.8B v3 remains listed (see above). The runs below are our own complete 0.2.1 runs.

Self-scored runs of the Decision Index 0.2.1 suite (kit 87d4650, the rebuilt suite is byte-identical to the lab's: rows b2b56d6f…, added rows 7429f3c9…), made at the request of the board maintainers (Space discussions #23 and #32).

Model Decision Index 0.2.1 Raw index Knowledge Language Retrieval Tools Arts Scored requests Median latency Scores
Jev-Style-0.8B-Decision-v3 @ 633729d8 16.64 35.98 5.6 10.6 26.2 35.4 7.3 150,317 (84 unsupported) 61 ms scores.json
Jev-Style-2B-Decision-v3 @ fbb9d676 34.83 50.14 22.3 29.0 51.6 52.8 15.9 150,317 (0 unsupported) 136 ms scores.json

Area columns are the chance-corrected area skills × 100. Latencies are in-process request wall times on the hardware below (not the maintainers' RTX PRO 6000 measurement).

How the v3 runs were made

  • Engine: engine/jev_engine_torch.py, a kit Engine over each model repo's own PyTorch runtime (jev_style_decision.py) at the pinned revision above, manifest-verified. float32 on CUDA; transformers 5.17.0, torch 2.11.0+cu128, flash-linear-attention 0.5.2 for the Gated DeltaNet layers (checked against the pure PyTorch reference on a 60-request sample: same top answer on 805/805 (2B) and 698/699 (0.8B) questions, max |Δp| ≤ 0.0025).
  • One fixed rendering and one global temperature for every benchmark; no prompt tuning, no option filtering, nothing truncated. Requests over a declared limit raise the kit's Unsupported and count as unanswered:
    • 0.8B: 25,600 tokens in total, and question + options + readout within 2,048 tokens; a choice question whose options do not fit is scored in option chunks with one softmax over all options, so only questions whose text alone exceeds the budget are unsupported.
    • 2B: 25,600 tokens in total; no separate question/options budget (numbered-option catalogue layout).
  • Run with python -m decision_index run --engine jev_engine_torch:JevStyleTorchEngine --option model=<0.8b|2b> --option revision=<rev> --edition 0.2.1 --compact. Colab reclaimed the VMs several times, so each run was resumed from its own results.jsonl, and the last part was split into contiguous row shards run in parallel and merged by run_id (every scored request appears exactly once; no error rows). environment.json lists every part.
  • jev-style-0.8b-decision-v3: 92,613 requests on an NVIDIA A100 (Colab), 58,146 on NVIDIA L4 (Colab).
  • jev-style-2b-decision-v3: 86,890 requests on an NVIDIA A100 (Colab), 63,869 on NVIDIA L4 (Colab).
  • The 2B's environment.json shows temperature: global (category=None) = None: the engine's provenance line reads a 0.8B-style attribute. The 2B runtime applies its one calibrated global temperature from readout_config.json (T = 0.8278650621) to every question; the engine file is published exactly as it was run.

Model cards: Jev-Style-2B-Decision-v4 (training pool: jev-style-2b-v4-training), Jev-Style-2B-Decision-v3, Jev-Style-0.8B-Decision-v3 (training pool: MacJev-0.8B-Decision-Data).

Models are Apache-2.0. Jev-Style 2B v4 was trained in part on soft labels distilled from TypeSafe Jev 1.14 via its public API (see its model card); the v3 models used no Jev outputs. Not affiliated with, endorsed by or connected to TypeSafe AI or Jev.
Downloads last month
142
Total file size:
135 MB

Collection including chaoliangUNSW/decision-index-results

Article mentioning chaoliangUNSW/decision-index-results