maxidl/comparia-french-maxx

Compar:IA-French-Maxx

Want to benchmaxx the Compar:IA arena leaderboard? This dataset contains optimized responses for real Compar:IA conversations. 6,575 assistant replies, each written to be the best possible answer to a real compar:IA conversation, paired with the weaker response(s) it beats. You can use them for SFT, preference optimization (e.g. DPO) or whatever else you are up to.

subset	rows	rejected	rejected_2	preference coverage	provenance
Arena - general	1,500	Harder of the two original arena replies after comparison with the optimal response; present on 1,485 rows	The other arena reply; present on 1,476 rows	895 strong, 590 medium, 15 none	arena:<response_id> -> ministere-culture/comparia-fr-arena.response_id
Arena - both_bad	2,400	Harder of two replies that the human voter rated both_bad	The other human-rejected reply	2,181 strong, 219 medium	arena:<response_id> -> ministere-culture/comparia-fr-arena.response_id
Reactions	2,675	Exact assistant message that received the user's negative reaction; present on 2,674 rows	Not applicable	2,283 strong, 246 medium, 145 weak, 1 none	reactions:<id> -> ministere-culture/comparia-reactions.id
Total	6,575	6,559 usable preference pairs	3,876 rows with a second negative	5,359 strong, 1,055 medium, 145 weak, 16 none	n/a

rejected is always the primary preference-training negative. For arena rows, rejected_2 retains the second original model response when available. For reaction rows, pref_strength reflects the independently verified reliability of the user's complaint: strong for valid, medium for partially valid, and weak for mistaken.

Filterable curation signals

Reaction rows retain the user's non-exclusive defect tags. They are nullable because they do not apply to arena rows:

column	tagged reaction rows	meaning
incorrect	1,185	The user marked the response as factually or substantively incorrect
superficial	918	The user marked the response as too shallow or incomplete
instructions_not_followed	817	The user marked an instruction-following failure

Conversation-shape fields apply to every row. conversation_turns counts user turns in the exact published messages context, is_multi_turn is true when that count exceeds one, and prompt_chars is the final user message's character count.

The source tags can overlap, and they describe the rejected response rather than the optimal response. For arena rows they are null, not false.

How the optimal responses were made

• Generated with an agentic pipeline using Le Chaton Fat, with independent cross-review and an absolute critic gate.
• arena: the optimal is blind pairwise-judged to beat both arena model responses, with compar:IA's style control applied so it wins on substance, not formatting.
• reactions: the optimal fixes the specific defect a user flagged, gated by an expert critic.
• Multi-turn arena conversations use the optimal earlier replies as history.

Columns

column	type	use
messages	list[{role, content}]	conversation history ending on the user turn. SFT context / preference prompt
response_optimal	string	the optimal reply. SFT target / preference chosen (plain, substance-first)
response_optimal_styled	string	same content, formatted (headers / bold / lists)
rejected	string | null	the response the optimal beat. Preference rejected (the hardest alternative)
rejected_2	string | null	a second weaker alternative when one exists (arena rows)
pref_strength	string | null	confidence that response_optimal ≻ rejected: strong / medium / weak; null when there is no negative
incorrect	bool | null	reaction-source defect tag; null for arena rows
superficial	bool | null	reaction-source defect tag; null for arena rows
instructions_not_followed	bool | null	reaction-source defect tag; null for arena rows
conversation_turns	int	number of user turns in messages
is_multi_turn	bool	whether conversation_turns > 1
prompt_chars	int	character count of the final user message
source	string	arena or reactions
id	string	provenance back to the source dataset

Use it

from datasets import load_dataset
ds = load_dataset("maxidl/comparia-french-maxx", split="train")

SFT - append the optimal reply as the final assistant turn:

sft = ds.map(lambda r: {"messages": r["messages"] + [{"role": "assistant", "content": r["response_optimal"]}]},
             remove_columns=ds.column_names)

Preference optimization (DPO / ORPO / KTO / …) - keep rows that have a negative (optionally only the strong ones):

pref = ds.filter(lambda r: r["rejected"] is not None)           # 6,559 pairs
# pref = pref.filter(lambda r: r["pref_strength"] == "strong")  # 5,359 cleanest pairs
pref = pref.map(lambda r: {"prompt": r["messages"],
                           "chosen": r["response_optimal"],
                           "rejected": r["rejected"]},
                remove_columns=ds.column_names)

Filter by failure mode or conversation shape:

instruction_failures = ds.filter(lambda r: r["instructions_not_followed"] is True)
multi_turn = ds.filter(lambda r: r["is_multi_turn"])
long_superficial = ds.filter(
    lambda r: r["superficial"] is True and r["prompt_chars"] >= 1_000
)

Swap response_optimal -> response_optimal_styled to train on the presentation-formatted variant, optimized for the non-style-controlled leaderboard.

Caveats

• Responses are model-generated; LLM-as-judge preferences approximate but do not equal human preference.
• weak/medium pairs are noisier. Filter on pref_strength for clean preference training.
• Source-data licensing and privacy terms of compar:IA apply (Etalab 2.0 / CC-BY-4.0, PII-filtered, unmoderated content retained).