Mvanypersele/luciole_RAG_formated

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

• HotpotQA: Yang et al. (2018), HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering, distributed under CC BY-SA 4.0.

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

• Trust-Align: Song et al. (2025), Measuring and Enhancing Trustworthiness of LLMs in RAG through Grounded Attributions and Learning to Refuse, data distributed under Apache 2.0. The GPT-4 targets are not included.
• ASQA: Stelmakh et al. (2022), ASQA: Factoid Questions Meet Long-Form Answers, distributed under Apache 2.0.
• QAMPARI: Amouyal et al. (2023), QAMPARI: A Benchmark for Open-domain Questions with Many Answers, distributed under CC0 1.0.
• Passages: Wikipedia (DPR 2018 dump, used by ALCE), CC BY-SA.

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

• HotpotQA-fr (abstracts): Mvanypersele/hotpotqa-fr-abstracts, built on French Wikipedia in the format of HotpotQA (Yang et al., 2018), distributed under CC BY-SA 4.0.

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