Hugging Face's logo

Datasets:
IO-AIR
/
arenamax

Tasks:
Text Generation
Question Answering
Languages:
English
Size:
100K<n<1M
Tags:
preference
dpo
human-preference
arena
rlhf
dev-interaction
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
id: string
prompt: string
chosen: string
rejected: string
meta: struct<source: string, license: string, chosen_model: string, rejected_model: string, preference: st (... 106 chars omitted)
  child 0, source: string
  child 1, license: string
  child 2, chosen_model: string
  child 3, rejected_model: string
  child 4, preference: string
  child 5, len_chosen: int64
  child 6, len_rejected: int64
  child 7, len_ratio: double
  child 8, rejected_refusal: bool
  child 9, category: string
tags: list<item: string>
  child 0, item: string
pair_categories: struct<knowledge_reasoning: int64, writing: int64, truthful_denial: int64, dev_interaction: int64, v (... 18 chars omitted)
  child 0, knowledge_reasoning: int64
  child 1, writing: int64
  child 2, truthful_denial: int64
  child 3, dev_interaction: int64
  child 4, visual_pref: int64
excluded_on_license: list<item: string>
  child 0, item: string
version: string
quality: list<item: string>
  child 0, item: string
card_rev: string
configs: struct<sft: int64, preference: int64, preference_tight: int64>
  child 0, sft: int64
  child 1, preference: int64
  child 2, preference_tight: int64
sources: struct<lmarena-ai/arena-human-preference-55k: string, argilla/ultrafeedback-binarized-preferences-cl (... 71 chars omitted)
  child 0, lmarena-ai/arena-human-preference-55k: string
  child 1, argilla/ultrafeedback-binarized-preferences-cleaned: string
  child 2, NVIDIA/HelpSteer2: string
  child 3, OpenAssistant/oasst2: string
name: string
gpt2_total: int64
to
{'name': Value('string'), 'version': Value('string'), 'configs': {'sft': Value('int64'), 'preference': Value('int64'), 'preference_tight': Value('int64')}, 'gpt2_total': Value('int64'), 'pair_categories': {'knowledge_reasoning': Value('int64'), 'writing': Value('int64'), 'truthful_denial': Value('int64'), 'dev_interaction': Value('int64'), 'visual_pref': Value('int64')}, 'sources': {'lmarena-ai/arena-human-preference-55k': Value('string'), 'argilla/ultrafeedback-binarized-preferences-cleaned': Value('string'), 'NVIDIA/HelpSteer2': Value('string'), 'OpenAssistant/oasst2': Value('string')}, 'excluded_on_license': List(Value('string')), 'quality': List(Value('string')), 'card_rev': Value('string')}
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 2951, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, 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 547, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, 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
              id: string
              prompt: string
              chosen: string
              rejected: string
              meta: struct<source: string, license: string, chosen_model: string, rejected_model: string, preference: st (... 106 chars omitted)
                child 0, source: string
                child 1, license: string
                child 2, chosen_model: string
                child 3, rejected_model: string
                child 4, preference: string
                child 5, len_chosen: int64
                child 6, len_rejected: int64
                child 7, len_ratio: double
                child 8, rejected_refusal: bool
                child 9, category: string
              tags: list<item: string>
                child 0, item: string
              pair_categories: struct<knowledge_reasoning: int64, writing: int64, truthful_denial: int64, dev_interaction: int64, v (... 18 chars omitted)
                child 0, knowledge_reasoning: int64
                child 1, writing: int64
                child 2, truthful_denial: int64
                child 3, dev_interaction: int64
                child 4, visual_pref: int64
              excluded_on_license: list<item: string>
                child 0, item: string
              version: string
              quality: list<item: string>
                child 0, item: string
              card_rev: string
              configs: struct<sft: int64, preference: int64, preference_tight: int64>
                child 0, sft: int64
                child 1, preference: int64
                child 2, preference_tight: int64
              sources: struct<lmarena-ai/arena-human-preference-55k: string, argilla/ultrafeedback-binarized-preferences-cl (... 71 chars omitted)
                child 0, lmarena-ai/arena-human-preference-55k: string
                child 1, argilla/ultrafeedback-binarized-preferences-cleaned: string
                child 2, NVIDIA/HelpSteer2: string
                child 3, OpenAssistant/oasst2: string
              name: string
              gpt2_total: int64
              to
              {'name': Value('string'), 'version': Value('string'), 'configs': {'sft': Value('int64'), 'preference': Value('int64'), 'preference_tight': Value('int64')}, 'gpt2_total': Value('int64'), 'pair_categories': {'knowledge_reasoning': Value('int64'), 'writing': Value('int64'), 'truthful_denial': Value('int64'), 'dev_interaction': Value('int64'), 'visual_pref': Value('int64')}, 'sources': {'lmarena-ai/arena-human-preference-55k': Value('string'), 'argilla/ultrafeedback-binarized-preferences-cleaned': Value('string'), 'NVIDIA/HelpSteer2': Value('string'), 'OpenAssistant/oasst2': Value('string')}, 'excluded_on_license': List(Value('string')), 'quality': List(Value('string')), 'card_rev': 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.

ArenaMax · v3

Human-preferred outputs, remixed. Arena winners + rating-best-of-N, expanded in v3 with two more human-annotated sources and a category taxonomy that teaches the three things raw preference pairs miss:

category pairs what it teaches
knowledge_reasoning 80,246 correct, well-structured answers win
writing 12,698 human taste in prose
dev_interaction 6,412 how humans want a model to behave during development: code, debugging, tooling, workflow
visual_pref 1,073 human visual taste: frontend/UI/HTML/SVG/design prompts
truthful_denial 699 politely correcting a false premise or declining a bad approach — the winner pushes back, the loser plays along

Configs

file rows shape
sft.jsonl.gz 103,458 unified messages rows (assistant = preferred response)
preference.jsonl.gz 101,128 DPO pairs {prompt, chosen, rejected, meta}
preference_tight.jsonl.gz 65,696 pairs with len_ratio <= 1.5 — tight human wins, no verbosity blowouts

Real gpt2 total (sft config): 60,060,667.

Sources (license-pure only)

source license rows (sft) role
lmarena-ai/arena-human-preference-55k apache-2.0 ~34k decisive human arena winners
argilla/ultrafeedback-binarized-preferences-cleaned mit ~55k rating-best-of-N
NVIDIA/HelpSteer2 (v3) cc-by-4.0 ~3.4k NVIDIA-annotated pairings, decisive preference strength (abs >= 2) only
OpenAssistant/oasst2 (v3) apache-2.0 ~10.6k human review_result trees, rank-1 vs worst-ranked sibling

Excluded on license doctrine: arena-human-preference-100k (model-provider ToU), lmsys-chat-1m (research-restricted), chatbot_arena_conversations (gated).

Quality doctrine (v3)

  • decisive wins only: ties/both-bad dropped; HelpSteer2 pairs kept at abs(preference_strength) >= 2 (slight preferences are weak signal)
  • refusal winners dropped; non-English dropped (latin-ratio); response floor 120 chars
  • verbosity-skew guard: pairs where chosen > 2x rejected AND > 2400 chars dropped from pairs (a length win is not a quality win); preference_tight caps len_ratio <= 1.5
  • global exact-md5 + normalized near-dup dedup across ALL four sources
  • truthful_denial pairs are guarded: the correction is substantive (>220 chars), the prompt is not itself a fact-check request, and the rejected side shows no correction of its own (asymmetric correction = truthfulness gap, not quality gap)
  • every row carries meta.category and cat:<name> tags; pairs carry meta.len_ratio, winner/loser provenance, ratings where available

Usage 

from datasets import load_dataset
pref = load_dataset("IO-AIR/arenamax", data_files="preference.jsonl.gz", split="train")
dev  = [r for r in pref if r["meta"]["category"] == "dev_interaction"]
denial = [r for r in pref if r["meta"]["category"] == "truthful_denial"]

Related

  • IO-AIR/filterme-sft — the main 2B-token SFT ensemble (v1.7), strictly separate from this preference dataset.
Downloads last month
29
Total file size:
273 MB