Datasets:

Modalities:
Text
Formats:
parquet
ArXiv:
Dataset Viewer
Auto-converted to Parquet Duplicate
messages
listlengths
3
3
supporting_facts_titles
listlengths
0
2
chunks_total
int64
1
10
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "Anthony Ashley Cooper, 1st Earl of Shaftesbury", "John Locke" ]
6
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "L'ultimo giorno di Pompei", "The Last Day of Pompeii" ]
10
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "Louis Garrel", "The Dreamers (film)" ]
10
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "Oranges and Sunshine", "Emily Watson" ]
5
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "Westshore (Tampa)", "WestShore Plaza" ]
7
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "Hockey Night (1984 film)", "Rick Moranis" ]
10
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[]
3
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "Providence Hospital (Washington, D.C.)", "MedStar Washington Hospital Center" ]
8
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "Hagerstown Valley", "Great Appalachian Valley" ]
8
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "Sati Kazanova", "Fabrika" ]
9
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "Cross Florida Barge Canal", "Lake Worth Lagoon" ]
10
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[]
6
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "Garden State (film)", "Zach Braff" ]
10
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "Cameron's Line", "Taconic orogeny" ]
10
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "Don Simpson", "Flashdance" ]
7
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "Atsushi Takenouchi", "Butoh" ]
10
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[]
3
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "Lakeside railway station, England", "Lakeside and Haverthwaite Railway" ]
8
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "CityCenter", "Monte Carlo Resort and Casino" ]
10
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "Isaac Durnford", "Dino Dan" ]
8
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "Shame (Robbie Williams and Gary Barlow song)", "Blackbird (Beatles song)" ]
8
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "Combat boot", "Gore-Tex" ]
8
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "Lopezia", "Fatsia" ]
6
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "2013 Oceania Cup", "2014 Men's Hockey World Cup" ]
7
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[]
6
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "Farrukh Saleem", "Game theory" ]
10
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "Air Traffic", "Halestorm" ]
8
[ { "content": "You are an AI conversational assistant specialized in **information retrieval and synthesis**.\nYour goal is to provide **precise, reliable, and well-structured answers** using **only the retrieved documents** (`Context`).\nPrioritize **clarity, accuracy, and completeness** in your responses.\n\n#...
[ "2003–04 FC Bayern Munich season", "Roy Makaay" ]
9
End of preview. Expand in Data Studio

Luciole RAG (formatted)

Working repository of the retrieval-augmented generation (RAG) supervised fine-tuning configurations of the Luciole models. Configurations documented below have their own section; the other configurations are earlier iterations.

hotpotqa_cite

Dataset overview

Supervised fine-tuning data for retrieval-augmented generation, built to train the Luciole models for OpenRAG. Each example is a chat conversation in which the assistant answers a multi-hop question from retrieved Wikipedia paragraphs only, quotes the passages it relies on with a citation tag pointing to the source document, and declines to answer when the documents are not sufficient.

Split Rows Answerable Unanswerable
train 57,828 49,716 8,112 (14.0%)

There is a single split: held-out evaluation is meant to be done on separate benchmarks (for instance the official HotpotQA dev set), not on a slice of this data.

What to know about the data:

  • The prompt is the one OpenRAG sends at inference: same system instructions, same document layout ([Source N] blocks), and the target responses were generated under exactly that prompt.
  • Citations use a single format. Inline, the assistant wraps each quoted excerpt as <ref id="N">excerpt</ref>, where N is the number of the cited [Source N] document and the excerpt is copied from that document. In 3.1% of the examples, the system prompt instead asks for a final References section listing one <ref id="N"/> per cited document, with no markers in the text.
  • Every excerpt is verified against the cited document. An excerpt may be shortened with an ellipsis (A ... B): it is accepted when each fragment is verbatim in the cited document and the fragments appear in order. 21.2% of the examples hold at least one such excerpt.
  • Unanswerable examples are minimal pairs: each one is a copy of an answerable example (same question, same distractors) from which the gold evidence was removed, and both twins are always kept or dropped together. The refusal is the generator's own, adapted to the question; 29.0% of them quote, with inline <ref> tags, what the documents do cover before declining.

Data format

Each row has three fields:

  • messages: a 3-turn conversation (system, user, assistant).
  • supporting_facts_titles: titles of the gold documents needed to answer (empty for unanswerable examples), for citation precision and recall.
  • chunks_total: number of documents in the system prompt.

The system message holds the instructions, then the documents, numbered and separated as OpenRAG does (linagora/openrag 6c4a59c):

[Source 1]
* filename: <Wikipedia title>

[CHUNK_START]

<paragraph>

[CHUNK_END]----------

[Source 2]
...

The citation rule of the system prompt reads:

When quoting from the context, wrap each excerpt inline with <ref id="N">excerpt</ref>, where N is the number of the cited document (the N of its [Source N] marker) and excerpt is copied verbatim from that document. From a table, quote the row you use as it is written, e.g. <ref id="N">| Revenue | 1,234 | 1,100 |</ref>, and give any value you compute from it after the tag.

The table sentence is part of the prompt shared with the table QA data of the same mix; HotpotQA documents contain no tables.

Two instructions vary per example:

  • citation placement: inline (96.9%) or a final References section (3.1%);
  • insufficient context: a generic rule (state that the documents do not allow an answer and invite the user to rephrase or add documents), or a strict variant that dictates one of five exact refusal phrases. The strict variant is given to about 5% of the examples, answerable or not (402 of the 8,112 unanswerable, 2,411 of the 49,716 answerable): on an unanswerable example the target is the dictated phrase alone, on an answerable one it is the normal answer. The share of refusals is thus the same under both rules (14.3% vs 14.0%), so the strict rule does not signal a refusal. Before 2026-11-03 the strict variant was given to unanswerable examples only.

The user message is the question. Example of an answerable target:

Based on the provided documents, Robert Flores was replaced on the live afternoon
edition of "SportsCenter" by **John Buccigross** <ref id="5">Flores hosted the live
afternoon edition of "SportsCenter" from noon - 3 p.m. with Chris McKendry until
early September 2009, when he was replaced with John Buccigross.</ref>.

John Buccigross has been an anchor for ESPN since **1996** <ref id="4">John
Buccigross ( ) is an American sportscaster.  He has been an anchor for ESPN since
1996.</ref>.

Example of an unanswerable target, References placement:

Based on the provided documents, **Grandaddy** is the band that featured Jason Lytle.
[...]
However, the provided documents do not contain information about when **Cake** was
formed, nor do they explicitly state the formation year of Grandaddy. [...]

References:
<ref id="5"/>
<ref id="6"/>

Dataset construction

  1. Source. All 90,447 questions of the HotpotQA training set, distractor setting (2 gold paragraphs and 8 distractors per question). Up to 70% of the distractors are dropped at random and the documents are shuffled, which gives 5 to 10 documents per prompt for 99.3% of the examples (mean 7.6).

  2. Unanswerable twins. 22% of the questions get a twin without sufficient evidence: either both gold paragraphs are removed, or only one of them (the question stays half supported, the harder case). Removed paragraphs are replaced one for one by paragraphs about unrelated entities, so both twins show the same number of documents. In the final data, 60.8% of the refusals have no gold paragraph left and 39.2% keep one.

  3. Generation. Qwen3.6-35B-A3B without thinking mode, temperature 0.1, given exactly the training prompt (no gold answer). The citation placement is drawn at this step (References with probability 0.04) and stored, so that the generation prompt and the training prompt are the same. 110,314 responses.

  4. Verification. Mistral-Small-3.2-24B-Instruct-2506 as judge, plus deterministic checks. An answerable example is kept when:

    • the factual judge gives 5/5 (correct answer, all supporting facts used, sound reasoning; the judge sees the gold answer and the supporting fact titles);
    • the cited documents are exactly the gold documents (citation F1 = 1);
    • every <ref> points to an existing document and its excerpt is found in it (verbatim, or as in-order verbatim fragments around an ellipsis).

    An unanswerable example is kept when a refusal judge classifies the response as a refusal (it may describe what the documents cover, but must not answer), with the same check on any <ref> it contains. Twins are kept or dropped together. 58,303 of the 110,314 examples pass (52.9%).

  5. Formatting. The response is used as generated, with two edits: an ellipsis that skips no word of the source (a trailing ..., or fragments adjacent in the document) is removed from the excerpt; the target of the unanswerable strict-refusal examples is replaced by the dictated phrase, while answerable examples drawn for the strict rule keep their answer.

  6. Cleanup. 400 questions are dropped with their twin: 126 whose answer names a [Source N] it does not cite with a tag, and 274 whose prompt holds residual Wikipedia <ref name=...> markup from the source paragraphs, which collides with the citation tags. 57,828 rows remain.

The filter does not skew the question mix: bridge questions are 80.7% of the generated answerable examples and 81.2% of the kept ones (comparison 18.8%); HotpotQA levels in the kept data are easy 15.2%, medium 67.8%, hard 17.0%.

Statistics

Value
Documents per prompt 2 to 10, mean 7.6
Prompt + response length, Luciole tokenizer (5,000 sampled rows) median 1,717, p99 2,993, max 4,342 tokens
Answerable responses, words median 80, p90 128
Refusals, words median 90, p90 180
<ref> tags per answerable response mean 2.13 (2 distinct documents in 99.4%)
Examples with an elided excerpt (A ... B) 21.2%

Known limitations

  • Responses are model-generated and model-judged. The judges and checks remove most errors, not all.
  • 9.1% of the responses also name the source in the text ("According to Source 4, ...") next to the tag.
  • The filter keeps only answers that cite exactly the gold documents; answers that were correct but cited one document out of two are not represented.

Source data and licenses

Since HotpotQA is released under a share-alike license, this dataset is distributed under CC BY-SA 4.0.

trust_align

Dataset overview

Supervised fine-tuning data for retrieval-augmented generation on questions with several answers spread over several documents: ambiguous questions with one answer per interpretation (ASQA) and list questions (QAMPARI). Each example is a chat conversation in which the assistant answers from retrieved Wikipedia passages only, quotes the passages it relies on with <ref id="N">excerpt</ref> tags, gives the part of the answer the documents support when they hold only part of it, and declines when they support none of it. The questions and their retrieved passages come from the Trust-Align training data; the responses are new.

Split Rows Answerable Unanswerable
train 1,737 1,457 280 (16.1%)
validation 183 157 26 (14.2%)

The validation split comes from the test split of the Trust-Align SFT data (questions of the ASQA and QAMPARI training sets held out by Trust-Align); the one question it shared with the train split was removed.

What to know about the data:

  • The system prompt, the document layout ([Source N] blocks) and the citation format are those of hotpotqa_cite; see that section for the full prompt.
  • Most answers are partial: in 75.5% of the answerable training examples, the documents hold only some of the expected answers, and the response gives those.
  • Answers cite more documents than in hotpotqa_cite: 2.81 distinct documents per answerable response on average, 56.3% cite three or more.
  • Unanswerable examples are not twins of answerable ones: they are other questions whose retrieved passages support none of the expected answers. 73.2% of the refusals quote what the documents do cover before declining.
  • An excerpt shortened with an ellipsis (A ... B) was split into consecutive tags <ref id="N">A</ref><ref id="N">B</ref>, each verbatim; the only ellipses left (9 examples) are part of the source text.

Data format

Same three fields as hotpotqa_cite: messages (system, user, assistant), supporting_facts_titles (the documents the Trust-Align target cited, empty for unanswerable examples) and chunks_total (number of documents in the system prompt). All examples use inline citations (no References section). The strict refusal instruction is used by 18 of the 280 training refusals, whose target is the dictated phrase alone, and by 59 of the 1,457 answerable training examples (9 of the 157 in validation), whose target is the normal answer; before 2026-11-03 it was given to refusals only.

Example of an answerable target (abridged):

The NCAA Division I Men's Basketball Tournament is broadcast across four networks:
**CBS, TNT, TBS, and truTV** <ref id="1">The coverage of the tournament is split
between CBS, TNT, TBS, and truTV.</ref><ref id="3">Through the agreement between CBS
and Turner, which began with the 2011 tournament, games are televised on CBS, TNT,
TBS and truTV.</ref>.

Dataset construction

  1. Source. The ASQA and QAMPARI rows of the Trust-Align SFT data (declare-lab/Trust-Score, Trust-Align/sft, commit 0b89ac0): a question, 5 Wikipedia passages retrieved for it, and a target written by GPT-4. ELI5 rows are left out. Gold answers are joined from ASQA and QAMPARI. Only the question and the passages are reused; the GPT-4 target serves to choose the reference: the supporting documents are the passages it cites, the reference answer is the gold answers it contains. Rows labelled unanswerable by Trust-Align are kept only when no gold answer string occurs in their passages, and sampled down to 20% of the rows.

  2. Documents. 4 BM25 distractors drawn from the passages of other questions (from rank 5, excluding passages that contain a gold answer) are added to the 5 retrieved passages, then up to 70% of the non-supporting documents are dropped at random and the documents are shuffled: 3 to 9 documents per prompt (mean 7.0).

  3. Generation. Qwen3.5-35B-A3B (FP8) without thinking mode, temperature 0.1, no gold answer in the prompt. 3,430 train and 365 validation responses. The prompt was an earlier version of the training prompt: same rules, documents numbered [N] filename: instead of [Source N] blocks, and without the table sentence. The documents were laid out again in the OpenRAG format afterwards (step 6).

  4. Verification. Qwen3.8-27B as judge (thinking mode), plus deterministic checks. An answerable example is kept when:

    • the factual judge gives 5/5 under a rubric for questions with several answers: every reference answer is given, each statement is supported by the documents, and further answers or cited documents are accepted when the documents support them;
    • every <ref> points to an existing document and its excerpt is verbatim in it, after the ellipsis split described above.

    There is no citation F1 requirement: the reference documents are the Trust-Align selection, not a gold set. An unanswerable example is kept when a refusal judge confirms the response is a refusal; 58% of the Trust-Align refusals were confirmed, the others being answerable from the passages.

  5. Formatting. The targets of the unanswerable strict-refusal examples are replaced by the dictated phrase; answerable examples drawn for the strict rule keep their answer.

  6. Layout and cleanup. The documents are laid out as in hotpotqa_cite. Responses that name a document by number outside a tag ("Document 3", "[5]"), which refer to the earlier numbering, are dropped (57), as are answerable responses without any <ref> tag (73).

Statistics

Value
Documents per prompt 3 to 9, mean 7.0
Prompt + response length, Luciole tokenizer (train) median 1,880, p99 2,863, max 3,979 tokens
Answerable responses, words median 136, p90 255
Refusals, words median 137, p90 226
<ref> tags per answerable response mean 4.31
Distinct cited documents per answerable response mean 2.81; 1: 12.6%, 3 or more: 56.3%
ASQA / QAMPARI (train) 1,259 / 478

Known limitations

  • Responses are model-generated and model-judged, by two models of the same family.
  • The filter keeps fewer questions with many answers: among answerable responses, 63% pass when the documents hold one expected answer, 41% when they hold five or more.
  • The reference answer is restricted to the gold answers present in the GPT-4 target, which may miss answers present in the documents; the rubric accepts such answers when they are supported.
  • Four QAMPARI questions appear twice in the train split, with different documents.

Source data and licenses

Since the passages are under a share-alike license, this configuration is distributed under CC BY-SA 4.0.

hotpotqa_fr_abstracts_v2

Dataset overview

French counterpart of hotpotqa_cite: multi-hop questions written on French Wikipedia, answered from retrieved paragraphs only, with the same OpenRAG prompt layout, the same <ref id="N"> citation format and the same verification. This is the French data of the Luciole RAG full-SFT mix of 2026-11-03. It replaces hotpotqa_fr_abstracts, whose prompts had French system instructions and whose refusals nearly all opened with the same sentence (see step 7 of the construction).

Split Rows Answerable Unanswerable
train 24,782 20,418 4,364 (17.6%)

What to know about the data:

  • The system instructions are the English OpenRAG prompt of hotpotqa_cite; the documents, the question and the response are in French (the prompt asks to answer in the language of the user's query). The same instructions were used at generation time.
  • Citations: inline <ref id="N">excerpt</ref>, or, in 3.3% of the examples, a final References section with one <ref id="N"/> per cited document. Every excerpt is verified against the cited document; 9.8% of the examples hold an excerpt shortened with an ellipsis whose fragments are verbatim and in order.
  • Unanswerable examples are minimal pairs: a copy of an answerable example with both gold paragraphs removed; twins are always kept or dropped together.
  • The opening of each refusal written under the generic rule was rewritten by a deterministic script (step 7). The body of the refusal and its <ref> tags are the generator's.

Data format

Same three fields and same document layout as hotpotqa_cite (messages, supporting_facts_titles, chunks_total).

The insufficient-context instruction is the generic rule for 94.8% of the examples and the strict variant (one of five dictated exact refusal phrases) for 5.2%, answerable or not (1,044 answerable, 244 unanswerable). The share of refusals is close under both rules (18.9% vs 17.5%), so the strict rule does not signal a refusal. Under the strict rule, the target of an unanswerable example is the dictated phrase alone.

Example of an answerable target:

Gościnowo est un village situé dans la **gmina de Skwierzyna** <ref id="8">Gościnowo
est un village polonais de la gmina de Skwierzyna dans le powiat de Międzyrzecz de la
voïvodie de Lubusz dans l'ouest de la Pologne.</ref>.

La superficie de la gmina de Skwierzyna est de **285,44 km²** <ref id="7">La gmina
couvre une superficie de 285,44 kilomètres carrés pour une population de 12 335
habitants en 2016.</ref>.

Example of an unanswerable target (opening rewritten, verdict kept in the body):

D'après le contexte fourni, je constate que les documents fournis ne contiennent aucune
information concernant le Championnat sud-américain de football de 1937, ni le nom du
buteur argentin de cette finale, ni sa ville de naissance.

Par conséquent, il est impossible de répondre à cette question en se basant uniquement
sur les documents récupérés.

Les documents récupérés ne me permettent pas de répondre à votre question. Pourriez-vous
la reformuler ou ajouter des documents pertinents ?

Dataset construction

  1. Source. The train split of the filtered_qwen_v2 configuration of Mvanypersele/hotpotqa-fr-abstracts (revision e0427e5): 29,080 multi-hop questions built on French Wikipedia abstracts, each with 2 gold abstracts and 8 distractors. Distractors are dropped at random and the documents shuffled, which gives 5 to 10 documents per prompt (mean 7.7). 4 questions whose documents exceed the 8,192-token budget are dropped.
  2. Unanswerable twins. 22% of the questions (6,396) get a twin from which both gold abstracts are removed.
  3. Generation. Qwen3.6-35B-A3B without thinking mode, temperature 0.1, given exactly the training prompt (no gold answer). 35,472 responses.
  4. Verification. Mistral-Small-3.2-24B-Instruct-2506 as judge, plus deterministic checks. An answerable example is kept when the factual judge gives 5/5, a French-quality judge gives 5/5 (fluent, correct French), the cited documents are exactly the gold documents (citation F1 = 1) and every <ref> excerpt is found in the cited document. An unanswerable example is kept when the refusal judge classifies the response as a refusal (96.8% of the twins), with the same <ref> check. Twins are kept or dropped together: 24,792 of 35,472 examples pass (69.9%); 1,940 pairs are lost because the answerable side failed.
  5. Formatting. An ellipsis that skips no word of the source is removed from the excerpt; the target of the strict-rule unanswerable examples is replaced by the dictated phrase.
  6. Cleanup. 1 question whose response is in English is dropped with its twin, and 8 rows whose answer appears only inside a <ref> tag (6 answerable rows, 2 twins). 24,782 rows remain.
  7. Refusal openings. In hotpotqa_fr_abstracts, 89.5% of the refusals started with the same sentence, the verdict first, almost word for word the strict refusal phrase of the French evaluation; a model trained on it refused answerable French questions. Here every refusal written under the generic rule goes through a deterministic rewrite: the verdict sentence is moved to the end, varied, or the response opens with a phrase that also opens answerable responses ("D'après les documents fournis", "Pour répondre à cette question", "Selon les documents fournis"). <ref> tags are never modified. 3,485 of the 4,364 refusals were changed; strict-rule refusals are untouched.

How refusals start (commitment = mean probability of a refusal given the first words of the response, measured on the 10% refusal mix):

hotpotqa_fr_abstracts this config hotpotqa_cite (English)
Most frequent first sentence 89.5% 4.6% 3.5%
Opening with the strict refusal phrase 95.8% 0.1% 0%
Commitment after the first word 0.89 0.39 0.12
Commitment after the first two words 0.93 0.54 0.35

Statistics

Value
Documents per prompt 5 to 10, mean 7.7
Answerable responses, words median 75, p90 126
Refusals, words median 82, p90 131
<ref> tags per answerable response mean 2.10 (2 distinct documents in 99.98%)
Examples with an elided excerpt (A ... B) 9.8%
Refusals quoting with <ref> 9.0%

Known limitations

  • Responses are model-generated and model-judged. The judges and checks remove most errors, not all.
  • All questions are bridge questions labelled hard by the source, and no refusal keeps one gold paragraph (no half-supported case, unlike hotpotqa_cite).
  • The citation filter keeps only answers that cite exactly the gold documents; 23% of the answerable responses fail it.
  • Artifacts of the opening rewrite: 379 refusals (8.7%) repeat the subject in the first sentence ("D'après les documents fournis, je constate que les documents fournis ne contiennent…"), and 15 have their closing sentence placed after the References header.
  • 46% of the refusals name a document by its number outside a tag ("Le document [Source 3] mentionne…"; 31% in hotpotqa_cite), and 28 answerable responses name a document they do not cite with a tag.
  • The opening of a response still predicts a refusal more than in English (commitment 0.39 vs 0.12 after the first word).

Source data and licenses

Since the source is released under a share-alike license, this configuration is distributed under CC BY-SA 4.0.

Thinking configurations (*_think)

Dataset overview

Responses regenerated from scratch with the thinking mode of the generator: the assistant message holds the model's own reasoning in a <think> block, followed by the final answer. The reasoning was not added afterwards to existing answers: the answer is the one the model wrote after that reasoning. Each configuration follows the recipe of its parent, on a sample of its questions.

Config Parent Generated Kept by the filter Final rows Answerable Unanswerable
hotpotqa_cite_think hotpotqa_cite 3,415 1,795 1,679 1,469 210 (12.5%)
hotpotqa_fr_abstracts_think hotpotqa_fr_abstracts_v2 1,952 1,164 1,056 870 186 (17.6%)
tatqa_cite_think tatqa_cite 1,586 690 651 580 71 (10.9%)
trust_align_think trust_align 600 338 327 272 55 (16.8%)
total 7,553 3,987 3,713 3,191 522 (14.1%)

What to know about the data:

  • Format. The prompt is that of the parent configurations ([Source N] documents, <ref> citations). The reasoning refers to the documents by these labels: 99.2% of the traces name a "Source N", so using another document layout requires rewriting the traces too.
  • Chat template. Check that your chat template keeps the <think> block in assistant messages; some templates strip it.
  • Refusal rule. Every example uses the generic insufficient-context instruction; the strict variant, which dictates the refusal phrase, is not used here.
  • Overlap with the parent configurations. Part of the questions also appear, answered without thinking, in the parent: 71.4% for hotpotqa_cite_think, 75.6% for hotpotqa_fr_abstracts_think, 47.6% for tatqa_cite_think, 67.9% for trust_align_think. Exclude them from the parent when combining the two, so that a question is not seen both with and without reasoning.

Data format

The fields of the parent configurations, messages, supporting_facts_titles and chunks_total, plus:

  • source: the parent configuration.
  • is_unanswerable: true when the documents are not sufficient and the target is a refusal.

The assistant message has the form:

<think>
{reasoning}
</think>

{answer}

Dataset construction

  1. Questions. 2,800 questions for HotpotQA, 1,600 for HotpotQA-fr, 1,300 for TAT-QA and 600 for Trust-Align (seed 42), with the document selection of the parent configurations. 22% of the HotpotQA, HotpotQA-fr and TAT-QA questions get an unanswerable twin; in half of them the evidence is only partly removed (also for HotpotQA-fr, whose parent removes both gold abstracts).
  2. Generation. Qwen3.6-35B-A3B with thinking mode, temperature 0.6, at most 12,000 tokens, given exactly the training prompt (no gold answer). A response whose reasoning is missing or cut off counts as failed and is retried, up to three attempts.
  3. Verification. Mistral-Small-3.2-24B-Instruct-2506 as judge for all four sources, plus deterministic checks. An answerable example is kept when the factual judge gives 5/5, the cited documents are exactly the gold documents (citation F1 = 1, not applied to trust_align_think), every <ref> excerpt is found in the cited document and, for French, a French-quality judge gives 5/5. An unanswerable example is kept when the refusal judge classifies the response as a refusal, with the same <ref> check. Twins are kept or dropped together. The refusal judge confirmed 64.9% of the unanswerable responses on HotpotQA, 95.5% on HotpotQA-fr, 40.6% on TAT-QA and 51.9% on Trust-Align.
  4. Formatting. In the answer, backticks around <ref> tags are removed, as well as ellipses that skip no word of the source. The reasoning is kept exactly as generated.
  5. Cleanup. Answerable rows whose answer appears only inside a <ref> tag are dropped with their twin (116 rows for hotpotqa_cite_think, 108 for hotpotqa_fr_abstracts_think, 39 for tatqa_cite_think, 11 for trust_align_think).

Statistics

Lengths in Luciole tokenizer tokens.

hotpotqa_cite_think hotpotqa_fr_abstracts_think tatqa_cite_think trust_align_think
Documents per prompt, mean 7.7 7.7 9.6 7.0
Prompt + response, median 2,627 3,145 3,573 2,789
Prompt + response, p99 7,031 5,902 6,449 5,210
Prompt + response, max 11,126 9,314 8,094 10,279
Rows over 8,192 tokens 12 2 0 1
Reasoning, median (answerable) 827 1,280 1,129 1,050
Reasoning, median (unanswerable) 1,060 1,020 878 914
Answer, median (answerable) 115 98 112 157
Answerable answers with a final References section 4.0% 7.7% 3.6% 3.7%

Known limitations

  • Responses are model-generated and model-judged. The judges read the answer only, not the reasoning: an error in a trace that does not reach the answer is not detected.
  • Traces are kept as generated, including their defects. 78.9% to 96.3% of them end with process leftovers ("✅", "(Done.)", "Proceeds", "Output matches"), and 59.3% to 74.6% draft the answer in full before it is repeated after </think>.
  • The reasoning is mostly in English, also on French examples: 84.9% of the traces of hotpotqa_fr_abstracts_think open with "Here's a thinking process:".
  • French refusals were not rewritten as in hotpotqa_fr_abstracts_v2, since a rewritten answer could contradict its own reasoning: 69.9% of them start with the word "Les", and the first word of a response predicts a refusal more strongly (commitment 0.74, against 0.39 in hotpotqa_fr_abstracts_v2).
  • The filter keeps only the refusals the judge confirmed, which is a strong selection on TAT-QA and Trust-Align (step 3).
  • The configurations are small; tatqa_cite_think and trust_align_think hold fewer than 700 rows each.

Source data and licenses

Sources and licenses are those of the parent configurations.

sft_mix_20261109_xml_think_rows

The rows with reasoning selected for an SFT mix that uses the XML document layout. They come from the four *_think configurations above, with the prompt and the reasoning converted to that layout. The answers follow the same conversion, and the <ref> tags are unchanged.

source Rows Answerable Unanswerable
hotpotqa_cite 1,497 1,347 150
hotpotqa_fr_abstracts 859 773 86
tatqa_cite 631 568 63
trust_align 298 268 30
total 3,285 2,956 329
  • Layout. The documents are given as <documents><document id="N" filename="...">...</document></documents>, and N in <ref id="N"> is the id of the cited document. In the reasoning, [Source N] becomes document N and the quoted citation rule takes the XML wording.
  • Selection. From the 3,713 rows of the *_think configurations: 49 rows dropped because their reasoning keeps a marker of the former layout that cannot be rewritten safely, 14 over 8,160 tokens, 209 whose question is in the validation or reserve sets of the mix, and 156 to bring refusals to 10% of the rows.
  • Fields. messages, source (parent configuration) and is_unanswerable.

sft_mix_20261108_xml_explicit

Dataset overview

The RAG mix built to be added to a global supervised fine-tuning mix of the Luciole models: 75,270 training rows in the XML document layout. No response was generated for it: it is drawn from the four source configurations below and converted without any LLM.

source Rows Answerable Unanswerable Share of the mix
hotpotqa_cite 48,942 44,084 4,858 (9.9%) 65.0%
hotpotqa_fr_abstracts_v2 20,199 18,175 2,024 (10.0%) 26.8%
tatqa_cite 4,513 4,066 447 (9.9%) 6.0%
trust_align 1,616 1,454 162 (10.0%) 2.1%
train 75,270 67,779 7,491 (10.0%)
validation 1,003 895 108 (10.8%)

What to know about the data:

  • Layout. The documents are given as <documents><document id="N" filename="...">...</document></documents>, with an optional <context> element when a chunk has a summary. Citations are <ref id="N">excerpt</ref> where N is the id of the cited document; 92.4% of the training rows hold at least one.
  • Refusal rule. 71,544 training rows use the generic instruction (state that the retrieved documents do not allow an answer and invite the user to rephrase the query or add relevant documents); the other 3,726 (5.0%) dictate an exact refusal phrase.
  • No duplicates. Each row appears once, and no question is shared between train and validation. Answerable and unanswerable twins share a question, so the 75,270 rows hold 67,916 distinct questions.
  • No reasoning. The assistant messages hold no <think> block.

Data format

  • messages: a 3-turn conversation (system, user, assistant).
  • source: the configuration the row comes from.
  • is_unanswerable: true when the documents are not sufficient and the target is a refusal.

Dataset construction

  1. Sources. The four source configurations of sft_mix_20261103 (OpenLLM-France/Luciole_RAG), after the same cleanup. Each source is drawn with 10% unanswerable rows.
  2. Sampling. Compared with sft_mix_20261103, more of the English and French rows are used and TAT-QA and Trust-Align are not repeated. 8,201 rows are held out for later preference training and are not published here; their questions appear in neither split.
  3. Conversion. The [Source N] prompts are rewritten to the XML layout, with the document numbers unchanged. In the answers, mentions of a source outside the <ref> tags become document N; the <ref> tags are not modified.
  4. Validation. 1% of the question groups of HotpotQA, HotpotQA-fr and TAT-QA, plus the Trust-Align validation split, held apart from training with both twins. All 1,003 of its rows have their question in the sft_mix_20261103 validation split.
  5. Cleanup. Without any LLM, leftovers of the mention rewrite are repaired outside the <ref> tags, which stay unchanged (247 training rows: doubled parentheses, markers between backticks, series of markers, wrong French gender). Rows are removed when a citation excerpt is absent from its document, or when the answer shows the model deliberating ("Wait, let me re-read..."): 209 from train and 5 from validation.

Statistics

Lengths in Luciole tokenizer tokens, system, user and assistant messages together.

train validation
Median 1,701 1,733
p99 3,624 3,778
Max 7,256 4,819

Known limitations

  • No model had been trained on this exact mix when it was published, so its effect on evaluation results is not known.
  • TAT-QA, the only source of tables and financial reports, weighs 6.0% of the mix.
  • Responses are model-generated and model-judged, with the limitations listed in the sections of the source configurations.

Source data and licenses

Sources and licenses are those of the source configurations.

Downloads last month
785

Papers for Mvanypersele/luciole_RAG_formated