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
overall: struct<n: int64, plan_exact_match_rate: double, template_id_accuracy: double, fallback_rate: double, (... 86 chars omitted)
  child 0, n: int64
  child 1, plan_exact_match_rate: double
  child 2, template_id_accuracy: double
  child 3, fallback_rate: double
  child 4, mean_superfluous_steps: double
  child 5, mean_missing_steps: double
  child 6, misses: list<item: null>
      child 0, item: null
by_split: struct<dev: struct<n: int64, plan_exact_match_rate: double, template_id_accuracy: double, fallback_r (... 99 chars omitted)
  child 0, dev: struct<n: int64, plan_exact_match_rate: double, template_id_accuracy: double, fallback_rate: double, (... 86 chars omitted)
      child 0, n: int64
      child 1, plan_exact_match_rate: double
      child 2, template_id_accuracy: double
      child 3, fallback_rate: double
      child 4, mean_superfluous_steps: double
      child 5, mean_missing_steps: double
      child 6, misses: list<item: null>
          child 0, item: null
by_family: struct<extract_classify_write: struct<n: int64, plan_exact_match_rate: double, template_id_accuracy: (... 960 chars omitted)
  child 0, extract_classify_write: struct<n: int64, plan_exact_match_rate: double, template_id_accuracy: double, fallback_rate: double, (... 86 chars omitted)
      child 0, n: int64
      child 1, plan_exact_match_rate: double
      child 2, template_id_accuracy: double
      child 3, fallback_rate: double
      child 4, mean_superfluous_steps: double
      child 5, mea
...
    child 6, misses: list<item: null>
              child 0, item: null
      child 3, translate_knowledge_write: struct<n: int64, plan_exact_match_rate: double, template_id_accuracy: double, fallback_rate: double, (... 86 chars omitted)
          child 0, n: int64
          child 1, plan_exact_match_rate: double
          child 2, template_id_accuracy: double
          child 3, fallback_rate: double
          child 4, mean_superfluous_steps: double
          child 5, mean_missing_steps: double
          child 6, misses: list<item: null>
              child 0, item: null
      child 4, translate_summarize_format: struct<n: int64, plan_exact_match_rate: double, template_id_accuracy: double, fallback_rate: double, (... 86 chars omitted)
          child 0, n: int64
          child 1, plan_exact_match_rate: double
          child 2, template_id_accuracy: double
          child 3, fallback_rate: double
          child 4, mean_superfluous_steps: double
          child 5, mean_missing_steps: double
          child 6, misses: list<item: null>
              child 0, item: null
  child 3, error_counts: struct<>
live: struct<>
created_utc: timestamp[s]
decision: struct<frozen_profile: string, open_test: bool, distillation_warranted: bool, distillation_reason: s (... 35 chars omitted)
  child 0, frozen_profile: string
  child 1, open_test: bool
  child 2, distillation_warranted: bool
  child 3, distillation_reason: string
  child 4, reasons: list<item: string>
      child 0, item: string
to
{'created_utc': Value('timestamp[s]'), 'split': Value('string'), 'offline_template': {'overall': {'n': Value('int64'), 'plan_exact_match_rate': Value('float64'), 'template_id_accuracy': Value('float64'), 'fallback_rate': Value('float64'), 'mean_superfluous_steps': Value('float64'), 'mean_missing_steps': Value('float64'), 'misses': List(Value('null'))}, 'by_split': {'dev': {'n': Value('int64'), 'plan_exact_match_rate': Value('float64'), 'template_id_accuracy': Value('float64'), 'fallback_rate': Value('float64'), 'mean_superfluous_steps': Value('float64'), 'mean_missing_steps': Value('float64'), 'misses': List(Value('null'))}}, 'by_family': {'extract_classify_write': {'n': Value('int64'), 'plan_exact_match_rate': Value('float64'), 'template_id_accuracy': Value('float64'), 'fallback_rate': Value('float64'), 'mean_superfluous_steps': Value('float64'), 'mean_missing_steps': Value('float64'), 'misses': List(Value('null'))}, 'knowledge_explain': {'n': Value('int64'), 'plan_exact_match_rate': Value('float64'), 'template_id_accuracy': Value('float64'), 'fallback_rate': Value('float64'), 'mean_superfluous_steps': Value('float64'), 'mean_missing_steps': Value('float64'), 'misses': List(Value('null'))}, 'translate_code': {'n': Value('int64'), 'plan_exact_match_rate': Value('float64'), 'template_id_accuracy': Value('float64'), 'fallback_rate': Value('float64'), 'mean_superfluous_steps': Value('float64'), 'mean_missing_steps': Value('float64'), 'misses': List(Value('null'))}, 'translate_knowledge_write': {'n': Value('int64'), 'plan_exact_match_rate': Value('float64'), 'template_id_accuracy': Value('float64'), 'fallback_rate': Value('float64'), 'mean_superfluous_steps': Value('float64'), 'mean_missing_steps': Value('float64'), 'misses': List(Value('null'))}, 'translate_summarize_format': {'n': Value('int64'), 'plan_exact_match_rate': Value('float64'), 'template_id_accuracy': Value('float64'), 'fallback_rate': Value('float64'), 'mean_superfluous_steps': Value('float64'), 'mean_missing_steps': Value('float64'), 'misses': List(Value('null'))}}, 'error_counts': {}}, 'live': {}, 'decision': {'frozen_profile': Value('string'), 'open_test': Value('bool'), 'distillation_warranted': Value('bool'), 'distillation_reason': Value('string'), 'reasons': List(Value('string'))}, 'historical_reference': {'suite': Value('string'), 'hybrid_v3_plan_exact_match': Value('float64'), 'hybrid_v3_final': Value('float64'), 'oracle_final': Value('float64'), 'rules_direct_final': Value('float64')}}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, 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 127, 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 483, 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 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, 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 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, 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 343, 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 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              overall: struct<n: int64, plan_exact_match_rate: double, template_id_accuracy: double, fallback_rate: double, (... 86 chars omitted)
                child 0, n: int64
                child 1, plan_exact_match_rate: double
                child 2, template_id_accuracy: double
                child 3, fallback_rate: double
                child 4, mean_superfluous_steps: double
                child 5, mean_missing_steps: double
                child 6, misses: list<item: null>
                    child 0, item: null
              by_split: struct<dev: struct<n: int64, plan_exact_match_rate: double, template_id_accuracy: double, fallback_r (... 99 chars omitted)
                child 0, dev: struct<n: int64, plan_exact_match_rate: double, template_id_accuracy: double, fallback_rate: double, (... 86 chars omitted)
                    child 0, n: int64
                    child 1, plan_exact_match_rate: double
                    child 2, template_id_accuracy: double
                    child 3, fallback_rate: double
                    child 4, mean_superfluous_steps: double
                    child 5, mean_missing_steps: double
                    child 6, misses: list<item: null>
                        child 0, item: null
              by_family: struct<extract_classify_write: struct<n: int64, plan_exact_match_rate: double, template_id_accuracy: (... 960 chars omitted)
                child 0, extract_classify_write: struct<n: int64, plan_exact_match_rate: double, template_id_accuracy: double, fallback_rate: double, (... 86 chars omitted)
                    child 0, n: int64
                    child 1, plan_exact_match_rate: double
                    child 2, template_id_accuracy: double
                    child 3, fallback_rate: double
                    child 4, mean_superfluous_steps: double
                    child 5, mea
              ...
                  child 6, misses: list<item: null>
                            child 0, item: null
                    child 3, translate_knowledge_write: struct<n: int64, plan_exact_match_rate: double, template_id_accuracy: double, fallback_rate: double, (... 86 chars omitted)
                        child 0, n: int64
                        child 1, plan_exact_match_rate: double
                        child 2, template_id_accuracy: double
                        child 3, fallback_rate: double
                        child 4, mean_superfluous_steps: double
                        child 5, mean_missing_steps: double
                        child 6, misses: list<item: null>
                            child 0, item: null
                    child 4, translate_summarize_format: struct<n: int64, plan_exact_match_rate: double, template_id_accuracy: double, fallback_rate: double, (... 86 chars omitted)
                        child 0, n: int64
                        child 1, plan_exact_match_rate: double
                        child 2, template_id_accuracy: double
                        child 3, fallback_rate: double
                        child 4, mean_superfluous_steps: double
                        child 5, mean_missing_steps: double
                        child 6, misses: list<item: null>
                            child 0, item: null
                child 3, error_counts: struct<>
              live: struct<>
              created_utc: timestamp[s]
              decision: struct<frozen_profile: string, open_test: bool, distillation_warranted: bool, distillation_reason: s (... 35 chars omitted)
                child 0, frozen_profile: string
                child 1, open_test: bool
                child 2, distillation_warranted: bool
                child 3, distillation_reason: string
                child 4, reasons: list<item: string>
                    child 0, item: string
              to
              {'created_utc': Value('timestamp[s]'), 'split': Value('string'), 'offline_template': {'overall': {'n': Value('int64'), 'plan_exact_match_rate': Value('float64'), 'template_id_accuracy': Value('float64'), 'fallback_rate': Value('float64'), 'mean_superfluous_steps': Value('float64'), 'mean_missing_steps': Value('float64'), 'misses': List(Value('null'))}, 'by_split': {'dev': {'n': Value('int64'), 'plan_exact_match_rate': Value('float64'), 'template_id_accuracy': Value('float64'), 'fallback_rate': Value('float64'), 'mean_superfluous_steps': Value('float64'), 'mean_missing_steps': Value('float64'), 'misses': List(Value('null'))}}, 'by_family': {'extract_classify_write': {'n': Value('int64'), 'plan_exact_match_rate': Value('float64'), 'template_id_accuracy': Value('float64'), 'fallback_rate': Value('float64'), 'mean_superfluous_steps': Value('float64'), 'mean_missing_steps': Value('float64'), 'misses': List(Value('null'))}, 'knowledge_explain': {'n': Value('int64'), 'plan_exact_match_rate': Value('float64'), 'template_id_accuracy': Value('float64'), 'fallback_rate': Value('float64'), 'mean_superfluous_steps': Value('float64'), 'mean_missing_steps': Value('float64'), 'misses': List(Value('null'))}, 'translate_code': {'n': Value('int64'), 'plan_exact_match_rate': Value('float64'), 'template_id_accuracy': Value('float64'), 'fallback_rate': Value('float64'), 'mean_superfluous_steps': Value('float64'), 'mean_missing_steps': Value('float64'), 'misses': List(Value('null'))}, 'translate_knowledge_write': {'n': Value('int64'), 'plan_exact_match_rate': Value('float64'), 'template_id_accuracy': Value('float64'), 'fallback_rate': Value('float64'), 'mean_superfluous_steps': Value('float64'), 'mean_missing_steps': Value('float64'), 'misses': List(Value('null'))}, 'translate_summarize_format': {'n': Value('int64'), 'plan_exact_match_rate': Value('float64'), 'template_id_accuracy': Value('float64'), 'fallback_rate': Value('float64'), 'mean_superfluous_steps': Value('float64'), 'mean_missing_steps': Value('float64'), 'misses': List(Value('null'))}}, 'error_counts': {}}, 'live': {}, 'decision': {'frozen_profile': Value('string'), 'open_test': Value('bool'), 'distillation_warranted': Value('bool'), 'distillation_reason': Value('string'), 'reasons': List(Value('string'))}, 'historical_reference': {'suite': Value('string'), 'hybrid_v3_plan_exact_match': Value('float64'), 'hybrid_v3_final': Value('float64'), 'oracle_final': Value('float64'), 'rules_direct_final': Value('float64')}}
              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.

UA Specialist Router β€” evaluation progress

Score tables, routing stats, and run metadata from the diploma project MariaOnyshchuk/ua-llm-router: a rules-based router over open Ukrainian specialists (Mamay-4B, Lapa-12B, Aya Expanse 8B, Qwen-Coder).

This dataset is the progress log of pinned JSON summaries, not a dump of every generation.

What is included

Path Contents
progress_ledger.csv Flattened metric rows across weeks (best table for browsing)
week_*/scores*.json, scores_mean_sd_*.json Suite-level quality and latency
week_*/specialist_matrix*.json Per-bucket bake-off used to set routes
week_*/*.meta.json Decoding (T=0, seed=42), model ids, GPU-seconds
week_*/routing_stats*.json Traffic mix for router runs
week_*/scores_detail.jsonl Item-level scores (ids + scores, no prompts)
v6_solos/scores.json Frozen mixed_ua_v6 1-pass dedicated solos (five models)
v6_solos/*.meta.json Decoding, GPU-seconds, model ids for those runs
tables/v6_solos_1pass.csv Same numbers as a wide CSV
tables/v6_preview_from_screen.csv Screening-sliced preview β€” do not mix with v6_solos
tables/a2_planner_grid*.csv Planning-only prompt Γ— model screen on composite v1
tables/a2_agentcoma*.csv Planning-only AgentCoMa diagnostics (IDs and metrics; no prompts or generations)
week_9_baselines/ Mamay-12B alone on mixed_ua_v4_balanced (S2 baseline), 3 repeats: scores, item-level scores, run metadata
week_11_pack_fp8/ Rules v2 router with three specialists in online FP8 on one GPU, 3 repeats: scores, item-level scores, run metadata, nvidia-smi after load
week_12_composite_v2_*/ Composite v2 planning results (offline template selector on dev and test). fixture/scores.json is a code check, see the note below
analysis/planner_split_v6_hybrid.json How often the Mamay-4B hybrid planner splits single-skill mixed_ua_v6 requests: 400 sampled items, 100 per claim bucket, counts and Wilson 95% intervals, no prompts. A planner diagnostic, not a system comparison
analysis/bootstrap_comparisons.json Paired item-level bootstrap (95% intervals) for router vs Mamay-12B comparisons on mixed_ua_v4_balanced

Headline pinned numbers (do not mix suites):

  • Scorer note (29 Sep 2026): from 4 Sep to 29 Sep score_results.py scored numeric-fact knowledge references (know-002 = 1991, know-004 = 24) as multiple-choice indexes, so every model got 0 on those two items (-0.0104 overall on v4). Fixed on 29 Sep. Rescored with the fixed scorer and published here: week 9 (Mamay-12B), week 11 (FP8 pack) and v6_solos. The 0.848 router and the week 6 solo numbers were scored before the bug and are unaffected. The week 6 and week 7 rescoring files made inside the bug window are not published; nothing from them is cited.
  • Rules v2 + few-shot on mixed_ua_v4 (192Γ—3): 0.848
  • Mamay-12B alone on mixed_ua_v4 (192Γ—3, S2 baseline): 0.884 (rescored; 0.874 before the fix), p50 1352 ms, about 2.5 GPU-seconds per prompt. The router: 0.848, p50 633 ms, about 1.5 GPU-seconds per prompt, three resident GPUs. The 12B model is the better single-number quality result; the router is faster and cheaper per prompt.
  • Same router with the three specialists in online FP8 on one GPU (week 11): 0.852 (rescored; 0.842 before the fix), p50 723 ms, 45 668 MiB resident. Knowledge is 0.750, the same as the bf16 router; the earlier drop to 0.688 came from the scorer bug.
  • Paired item-level bootstrap, router (FP8 run) minus Mamay-12B: overall -0.032, 95% interval [-0.072, +0.008], not significant (unchanged by the rescoring: both systems gain the same two items). Instruct: -0.156, [-0.281, -0.031], significant. The FP8 run stands in for the 0.848 router because item-level scores for the bf16 router are not stored. analysis/bootstrap_comparisons.json
  • Same router on bitsandbytes 4-bit (week 8): 0.830
  • Composite 36Γ—3 (week 10): oracle 0.882, one-hop 0.812, hybrid planner 0.768
  • mixed_ua_v5 (+HumanEval) 0.796 is a harder suite, not a regression vs 0.848
  • Frozen mixed_ua_v6 (2414 items, 1 pass, T=0, seed=42): dedicated solos in v6_solos/ β€” Mamay-12B 0.695, Lapa 0.670, Aya 0.668, Mamay-4B 0.666, Qwen-7B 0.595. Do not cite the incomplete Mamay-12B screening JSONL. Do not mix screening-sliced scores with this table.
  • Composite v2 (200 items, train/dev/test): the constrained template selector matches the stored workflow on all 60 dev and all 60 test items (exact workflow match 1.000). This is a planning result only. Specialist execution on the test split has not been run yet.
  • week_12_composite_v2_test_bakeoff/fixture/scores.json shows final scores of 1.0 for the oracle and template systems. Those runs use fixture specialists that return the expected answer, so they check the scoring and orchestration code. They say nothing about model quality and must not be cited as results.
  • A2 planner tables are diagnostics, not end-task quality results. The AgentCoMa intent sequence is a project-defined diagnostic oracle, not benchmark gold.

What is not included (on purpose)

  • Raw generation JSONL (prompt / content) β€” those rows reprint ZNO-Eval, FLORES-200, UAlign, and HumanEval text.
  • UA-Code-Bench / Eolymp problem statements (not redistributable).
  • AgentCoMa prompts or translated derivatives (the source dataset is gated and does not publish a reusable dataset license).
  • Empty failed stubs and composite smoke folders.

Prompts live in the original sources and in the GitHub repo’s benchmarks/ samples. Cite those datasets if you rebuild the suites.

Decoding

Unless a *.meta.json says otherwise: temperature 0.0, seed 42. Week 1 wiring smoke is unpinned and marked superseded in the ledger.

Load

from datasets import load_dataset

ledger = load_dataset("MariaOnyshchuk/ua-llm-router-eval", data_files="progress_ledger.csv")
print(ledger["train"][0])

Replace USERNAME with the Hugging Face namespace that hosts this repo.

Citation

@misc{onyshchuk-ua-llm-router-eval,
  title  = {UA Specialist Router evaluation progress},
  author = {Onyshchuk, Maria},
  year   = {2026},
  url    = {https://huggingface.co/datasets/MariaOnyshchuk/ua-llm-router-eval}
}

Code and thesis notes: https://github.com/MariaOnyshchuk/ua-llm-router

Downloads last month
78