Datasets:

Hunter-Seeker
/

agentrewardbench-results

Error code:   StreamingRowsError
Exception:    ValueError
Message:      Expected object or value
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 290, in _generate_tables
                  pa_table = paj.read_json(
                      io.BytesIO(batch), read_options=paj.ReadOptions(block_size=block_size)
                  )
                File "pyarrow/_json.pyx", line 342, in pyarrow._json.read_json
                File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
                  return check_status(status)
                File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
                  raise convert_status(status)
              pyarrow.lib.ArrowInvalid: JSON parse error: Column() changed from object to number in row 0
              
              During handling of the above exception, another exception occurred:
              
              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 304, in _generate_tables
                  batch = json_encode_fields_in_json_lines(original_batch, json_field_paths)
                File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 111, in json_encode_fields_in_json_lines
                  examples = [ujson_loads(line) for line in original_batch.splitlines()]
                              ~~~~~~~~~~~^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 20, in ujson_loads
                  return pd.io.json.ujson_loads(*args, **kwargs)
                         ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
              ValueError: Expected object or value

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.

Hunter-Seeker on AgentRewardBench — results

Per-benchmark, per-agent y_human / y_judge arrays for the Hunter-Seeker submission to the AgentRewardBench leaderboard. This is the artifact the leaderboard's submission form asks for as a Results URL.

#	Judge	Precision	Recall	F1
1	Hunter-Seeker	86.0	27.1	41.2
2	Rule-based validator	83.8	55.9	67.1
3	WebJudge (o4-mini)	82.0	47.8	60.4

First of seventeen on precision, the metric the leaderboard reports — judging web-agent trajectories from run telemetry alone. No goal, no final answer, no page content, no screenshot, and no model call at judgment time.

Per-benchmark precision: VisualWebArena 89.3 (first), WorkArena 100.0 (tied first), WorkArena++ 85.7, WebArena 79.1, AssistantBench — abstained.

What is in the file

results.json follows the leaderboard's expected shape:

{ "hunter-seeker": { <benchmark> : { <agent> : { <category> : { y_human: [...], y_judge: [...] } } } } }

with <benchmark> in assistantbench | webarena | visualwebarena | workarena | workarena++ | all, <agent> the four GenericAgent variants plus all, and <category> the four annotation categories. 1,106 test-split trajectories, 295 human-labelled successes.

Every headline number recomputes from this file — verify.py in the project repository runs 54 such checks, including rank against a snapshot of the live leaderboard.

Method, briefly

Hunter-Seeker is the decision layer for AI agents: given a table with an entity column and a yes/no outcome, it scores every row and refuses when nothing clears the bar. Agents call it when they must decide whether to act, escalate, or refuse; every decision comes back signed.

Here it judges trajectory success from run telemetry — step counts, step-budget exhaustion, retries, token volumes, elapsed time, and the shape of the action sequence. What separates the classes is behavioural and stark: trajectories it calls successful average 5.2 steps, no repeated consecutive actions and 53 seconds; the ones it does not average 21.6 steps, 6.9 repeats and 310 seconds, with 61% exhausting the step budget.

Scores are out-of-fold under task-grouped 5-fold cross-validation, averaged over fifty independent fold seeds. All reported figures are the output of AgentRewardBench's own scripts/score_judgments.py, not a reimplementation — the functional baseline run through the same path reproduces the published row exactly (83.8 / 55.9 / 67.1).

AssistantBench is a refusal, not a miss

Zero positive predictions on those 108 trajectories, so the reported precision is 0. It is not blindness — within AssistantBench the model still ranks trajectories at AUC 0.811 — but with 8 successes in 108 (a 7.4% base rate) nothing reaches the confidence the decision rule requires, and forcing calls there yields roughly 20% precision. Returning nothing rather than a guess is the intended behaviour.

Honest limitations

The margin over the rule-based validator is +2.2 precision points, paired bootstrap 95% CI [−4.5, +9.0]; over WebJudge (o4-mini), +4.0. First place is an ordering, not a demonstrated separation.
Recall is 27.1 against the validator's 55.9 — a high-precision, low-coverage judge, answering on 93 of 1,106 trajectories.
We are not first on every benchmark: WebJudge (o4-mini) leads WorkArena++, WebArena and AssistantBench.
1,106 trajectories over 300 tasks; per-benchmark cells rest on 7 to 43 calls and are underpowered.

Hunter-Seeker on AgentRewardBench — results

What is in the file

Method, briefly

AssistantBench is a refusal, not a miss

Honest limitations

Links