The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
aops/aops_1959_IMO_Problems/Problem_1: struct<mrr@10: double, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, ndcg@10: double
child 2, recall@100: double
aops/aops_1971_AHSME_Problems/Problem_26: struct<mrr@10: double, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, ndcg@10: double
child 2, recall@100: double
aops/aops_2000_AMC_12_Problems/Problem_4: struct<mrr@10: double, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, ndcg@10: double
child 2, recall@100: double
aops/aops_2002_AIME_I_Problems/Problem_6: struct<mrr@10: double, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, ndcg@10: double
child 2, recall@100: double
aops/aops_2009_AMC_12A_Problems/Problem_17: struct<mrr@10: double, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, ndcg@10: double
child 2, recall@100: double
aops/aops_2013_AMC_12A_Problems/Problem_13: struct<mrr@10: double, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, ndcg@10: double
child 2, recall@100: double
aops/aops_2015_AMC_12A_Problems/Problem_22: struct<mrr@10: double, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, ndcg@10: double
child 2, recall@100: double
aops/aops_2017_AIME_II_Problems/Problem_3: struct<mrr@10: double, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, ndcg@10: double
child 2, recall@100: double
aops/aops_2019_A
...
child 1, n: int64
child 2, ndcg@10: double
child 3, recall@100: double
child 8, stackoverflow: struct<mrr@10: double, n: int64, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, n: int64
child 2, ndcg@10: double
child 3, recall@100: double
child 9, sustainable_living: struct<mrr@10: double, n: int64, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, n: int64
child 2, ndcg@10: double
child 3, recall@100: double
child 10, theoremqa_questions: struct<mrr@10: double, n: int64, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, n: int64
child 2, ndcg@10: double
child 3, recall@100: double
child 11, theoremqa_theorems: struct<mrr@10: double, n: int64, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, n: int64
child 2, ndcg@10: double
child 3, recall@100: double
child 11, retries_per_query: double
child 12, spend_mean: double
child 13, thinking_mean: double
child 14, tool_calls_mean: double
child 15, utilisation: double
child 16, wall_parallel_mean: double
child 17, wall_sequential_mean: double
n_shards: int64
seed: int64
split: string
system: string
budget_multiple: int64
to
{'budget_multiple': Value('int64'), 'config_id': Value('string'), 'n_shards': Value('int64'), 'seed': Value('int64'), 'split': Value('string'), 'subsets': {'full': {'avg': {'mrr@10': Value('float64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'cached_input_mean': Value('float64'), 'cap_hit_rate': Value('float64'), 'cap_mean': Value('float64'), 'failed_rate': Value('float64'), 'fallback_rate': Value('float64'), 'fallback_share': Value('float64'), 'llm_calls_mean': Value('float64'), 'n_queries': Value('int64'), 'overshoots': Value('int64'), 'per_domain': {'aops': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'biology': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'earth_science': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'economics': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'leetcode': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'pony': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'psychology': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'robotics': {'mrr@10': Value('float64'), 'n': Value('
...
'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'economics': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'leetcode': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'pony': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'psychology': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'robotics': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'stackoverflow': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'sustainable_living': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'theoremqa_questions': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'theoremqa_theorems': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}}, 'retries_per_query': Value('float64'), 'spend_mean': Value('float64'), 'thinking_mean': Value('float64'), 'tool_calls_mean': Value('float64'), 'utilisation': Value('float64'), 'wall_parallel_mean': Value('float64'), 'wall_sequential_mean': Value('float64')}}, 'system': 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
aops/aops_1959_IMO_Problems/Problem_1: struct<mrr@10: double, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, ndcg@10: double
child 2, recall@100: double
aops/aops_1971_AHSME_Problems/Problem_26: struct<mrr@10: double, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, ndcg@10: double
child 2, recall@100: double
aops/aops_2000_AMC_12_Problems/Problem_4: struct<mrr@10: double, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, ndcg@10: double
child 2, recall@100: double
aops/aops_2002_AIME_I_Problems/Problem_6: struct<mrr@10: double, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, ndcg@10: double
child 2, recall@100: double
aops/aops_2009_AMC_12A_Problems/Problem_17: struct<mrr@10: double, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, ndcg@10: double
child 2, recall@100: double
aops/aops_2013_AMC_12A_Problems/Problem_13: struct<mrr@10: double, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, ndcg@10: double
child 2, recall@100: double
aops/aops_2015_AMC_12A_Problems/Problem_22: struct<mrr@10: double, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, ndcg@10: double
child 2, recall@100: double
aops/aops_2017_AIME_II_Problems/Problem_3: struct<mrr@10: double, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, ndcg@10: double
child 2, recall@100: double
aops/aops_2019_A
...
child 1, n: int64
child 2, ndcg@10: double
child 3, recall@100: double
child 8, stackoverflow: struct<mrr@10: double, n: int64, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, n: int64
child 2, ndcg@10: double
child 3, recall@100: double
child 9, sustainable_living: struct<mrr@10: double, n: int64, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, n: int64
child 2, ndcg@10: double
child 3, recall@100: double
child 10, theoremqa_questions: struct<mrr@10: double, n: int64, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, n: int64
child 2, ndcg@10: double
child 3, recall@100: double
child 11, theoremqa_theorems: struct<mrr@10: double, n: int64, ndcg@10: double, recall@100: double>
child 0, mrr@10: double
child 1, n: int64
child 2, ndcg@10: double
child 3, recall@100: double
child 11, retries_per_query: double
child 12, spend_mean: double
child 13, thinking_mean: double
child 14, tool_calls_mean: double
child 15, utilisation: double
child 16, wall_parallel_mean: double
child 17, wall_sequential_mean: double
n_shards: int64
seed: int64
split: string
system: string
budget_multiple: int64
to
{'budget_multiple': Value('int64'), 'config_id': Value('string'), 'n_shards': Value('int64'), 'seed': Value('int64'), 'split': Value('string'), 'subsets': {'full': {'avg': {'mrr@10': Value('float64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'cached_input_mean': Value('float64'), 'cap_hit_rate': Value('float64'), 'cap_mean': Value('float64'), 'failed_rate': Value('float64'), 'fallback_rate': Value('float64'), 'fallback_share': Value('float64'), 'llm_calls_mean': Value('float64'), 'n_queries': Value('int64'), 'overshoots': Value('int64'), 'per_domain': {'aops': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'biology': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'earth_science': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'economics': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'leetcode': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'pony': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'psychology': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'robotics': {'mrr@10': Value('float64'), 'n': Value('
...
'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'economics': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'leetcode': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'pony': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'psychology': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'robotics': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'stackoverflow': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'sustainable_living': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'theoremqa_questions': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}, 'theoremqa_theorems': {'mrr@10': Value('float64'), 'n': Value('int64'), 'ndcg@10': Value('float64'), 'recall@100': Value('float64')}}, 'retries_per_query': Value('float64'), 'spend_mean': Value('float64'), 'thinking_mean': Value('float64'), 'tool_calls_mean': Value('float64'), 'utilisation': Value('float64'), 'wall_parallel_mean': Value('float64'), 'wall_sequential_mean': Value('float64')}}, 'system': Value('string')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Multi-agent retrieval under a token budget: run logs (full_b14)
Complete outputs of one preregistered experiment comparing single-agent and multi-agent LLM retrieval systems on BRIGHT under matched token budgets (1× and 4× a base budget T). Every LLM call, tool call, ranking and per-query score is included, so every table can be recomputed.
- Code:
github.com/kheesu/multi-agent-rir-expenditure. Phase 1 and 2 ran at the commit of tagphase1-freeze-b14(ea80bac); Phase 3 at tagphase3-freeze-b14(6849afa). - Preregistration, amendments and decision log:
docs/. - Status: results are effect sizes. The significance testing for Phase 4 has not been run yet.
Setup
| Model | Qwen/Qwen3-8B-FP8 @ 220b46e3, vLLM 0.33.0, bf16 KV cache, 2× H100 80GB |
| Larger model (ablation A5) | Qwen/Qwen3-32B-FP8 @ aa55da1e |
| Retrievers | BM25 and AQ-MedAI/Diver-Retriever-4B (dense), as agent tools |
| Data | BRIGHT @ 3066d29c, 12 domains; frozen dev/test split (seed 20261108) |
| Base budget T | 30,468 tokens per query (the p90 of S1's spend on dev, P1-5) |
| Main test set (E1) | 399-query domain-stratified subset of the test split, at 1×T and 4×T, 3 seeds (S0: 1 seed) |
Systems: S0 retrieve+rerank, S0-B budget-scaled S0, S1 ReAct, S1-F budget forcing, S2 parallel S1 + RRF, S3 parallel S1 + verifier, M1 orchestrator-workers, M2 planner-inspectors.
Main result (Table 1, nDCG@10, mean ± sd over seeds)
| System | 1×T | 4×T | spend 1× → 4× (tokens/query) |
|---|---|---|---|
| S0 retrieve+rerank | 0.305 | 11.8k | |
| S0-B budget-scaled S0 | 0.309 ± 0.007 | 0.308 ± 0.003 | 23.6k → 95.3k |
| S1 ReAct | 0.339 ± 0.002 | 0.341 ± 0.002 | 16.1k → 17.1k |
| S1-F budget forcing | 0.337 ± 0.010 | 0.345 ± 0.009 | 25.4k → 69.5k |
| S2 parallel S1 + RRF | 0.338 ± 0.004 | 0.341 ± 0.003 | 16.1k → 64.2k |
| S3 parallel S1 + verifier | 0.298 ± 0.003 | 0.308 ± 0.006 | 16.2k → 50.2k |
| M1 orchestrator-workers | 0.325 ± 0.007 † | 0.336 ± 0.010 | 19.0k → 35.5k |
| M2 planner-inspectors | 0.255 ± 0.011 | 0.270 ± 0.006 | 22.9k → 90.0k |
† Fallback-dominated (>50% of queries scored from a fallback ranking); reported but excluded from claims about the design.
Full tables (rates, the non-RETECO subset, per-domain results, costs) and the Phase 2 and 3 reports are in exhibits/.
Layout
runs/<run_id>/ one folder per run (449). Run id: <stage>_<system>_b<budget>_s<seed>_<hash>[_sh<i>of8]
config.json full config + provenance (git SHA, model revision, vLLM version, GPU, KV dtype)
queries.jsonl one line per query: final ranking, ledger spend, cap hit, fallback, call counts, wall clock
calls.jsonl.gz one line per LLM / tool call: role, prompt and output tokens, ledger charge, retrieved ids
metrics.json run-level nDCG@10, MRR@10, Recall@100
per_query_metrics.json the same per query
errors.jsonl only where a query hit an infrastructure error; every such query was rerun and is in queries.jsonl
pipeline/ stage records of the automated pipeline (p1_*.json, p2_*.json, p3_*.json), audits
frozen/ the Phase 1 freeze: decisions, the Phase 2 matrix, subset.json (the 399 query ids)
exhibits/ tables (md/csv/tex), figures, phase2_report.md, phase3/ (ablations, per_query_deltas.json)
eval/ P2-6 evaluation of E1: per-configuration results (configs/), summary.json,
archive_SHA256SUMS (checksums of the uncompressed E1 logs as archived in P2-11)
docs/ preregistration.md (with amendments), decisions.md, deviations.md, pipeline.md
code/ git bundle of the code repository (full history; tags phase1-freeze-b14, phase3-freeze-b14):
git clone code/multi-agent-rir-expenditure.bundle
ops/ how the run was operated on the pod: server/pipeline/supervisor logs (logs/*.gz), env.sh,
supervisor state and history (node/state.json), the Phase 3 orchestration scripts,
and the superseded n=99 P1-4 record (one query was lost to a client bug and rerun)
Run-id prefixes: P15 (T measurement), P14 (external NeMo-style reference), P1 (Phase 1 tuning on dev),
P21/P23 (stack check, throughput), E1 (main experiment, 8 shards per configuration),
A1–A6 (Phase 3 ablations, seed 1, on the subset), P34* (larger model: V1 check and its own T).
Ablations: A1 ranking step, A2 worker count, A3 retriever tools, A4 shared board, A5 model size, A6 prompt wording.
On this run A1, A2 and A4 are at 4× instead of the preregistered 16× (E1 had no 16× arm); see Amendment 2 in
docs/preregistration.md.
Deviations from the preregistration (all documented in docs/)
- E1 ran on the 399-query stratified subset, not on all 1,246 test queries (compute; confidence intervals about 1.8× wider).
- The P1-8 grid for M1 and M2 was N ∈ {2, 4} × orchestrator share 0.2.
- KV cache in bf16, not fp8.
- Phase 3 used 4× instead of 16×, as above.
Reading the logs
import gzip, json
calls = [json.loads(l) for l in gzip.open("runs/<run_id>/calls.jsonl.gz", "rt")]
queries = [json.loads(l) for l in open("runs/<run_id>/queries.jsonl")]
Not included: eval/raw and archive/ (3.6 GB each; byte-identical copies of the E1 runs in runs/), the BRIGHT corpus and the retrieval indexes (rebuilt from the pinned BRIGHT revision by the setup scripts).
License and attribution
Released under CC BY 4.0. The prompts in calls.jsonl.gz contain BRIGHT queries and documents
(Su et al., 2024, BRIGHT: A Realistic and Challenging Benchmark for Reasoning-Intensive Retrieval, CC BY 4.0),
and these files reveal answers for BRIGHT's test queries. The model outputs come from Qwen3 models (Apache 2.0).
- Downloads last month
- 2