Dataset Card for Selene

Selene is a French instruction-tuning dataset specialized in creative writing (short stories, scenarios, dialogue, worldbuilding). It is designed to train a from-scratch instruct model on top of a raw base model (no instruct checkpoint), so that instruction-following, French grounding and a literary voice are all installed by supervised fine-tuning.

The dataset is uncensored for fiction: by curation it contains no refusals and no moralizing, so a model trained on it follows fiction instructions — including dark or mature themes — within a fiction frame, instead of declining. "Uncensored" here means the model obeys rather than refuses; the author remains responsible for the content, and quality is governed by dataset curation, not by removing safety behaviors.

• Curated by: Chortan
• Language: French (fr)
• Format: ChatML-style conversations, JSON Lines
• License: mixed sources — non-commercial / research only (see Licensing Information)

Dataset Summary

Selene merges two complementary layers:

1. General instruction layer (French) — installs instruction-following and keeps French fluent. As of v2, it also includes reconstructed multi-turn oasst2 threads and three additional French instruction sources (see Dataset Composition).
2. Creative layer (French) — the specialization: public-domain literary prose reformatted into instruction/response pairs. As of v2, it also includes true multi-turn "continue" chains (the next turn is the actual, consecutive next segment of the same public-domain book) and long-form single-turn passages generated under verifiable constraints (point of view, length, tone, genre, character). Purely synthetic creative data remains a possible future extension.

A per-example layer field lets you rebalance the instruct/creative ratio at training time. The published file is the merged raw pool and is currently instruction-heavy (94.3 % instruct / 5.7 % creative by count — see Dataset Composition); the intended training mix, produced by the project's build script, targets roughly 55 % instruction / 45 % creative by keeping all creative examples and subsampling instruction.

v2 note for downstream consumers: most examples are still single-turn (user → assistant), but messages can now contain more than two turns for the multi-turn sources (oasst2_thread, french_pd_chain, wikisource_chain) — about 1.2 % of the pool. Code that assumes exactly one user/assistant pair per example should be updated to iterate over messages and mask loss on every non-assistant turn.

Languages

All examples are in French (fr). Non-French content from multilingual sources was removed by language identification.

Dataset Structure

Each line is one conversation:

{
  "id": "french_pd_books-1234",
  "messages": [
    {"role": "user", "content": "Écris une scène de retrouvailles tendue entre deux anciens amants."},
    {"role": "assistant", "content": "Le vestibule sentait la pluie et la cire..."}
  ],
  "source": "french_pd_books",
  "layer": "creatif",
  "lang": "fr",
  "needs_lid": false
}

A multi-turn example (v2, french_pd_chain) — the second user turn asks for a continuation and the second assistant turn is the true, consecutive next segment of the same source book:

{
  "id": "french_pd_chain-0042",
  "messages": [
    {"role": "user", "content": "Écris un passage où un mousquetaire rejoint un fort assiégé, à la troisième personne, environ 2000 mots."},
    {"role": "assistant", "content": "Le vent soufflait sur les remparts..."},
    {"role": "user", "content": "Continue le récit."},
    {"role": "assistant", "content": "Le lendemain, l'aube se leva sur..."}
  ],
  "source": "french_pd_chain",
  "layer": "creatif",
  "lang": "fr",
  "needs_lid": false
}

Fields (schema unchanged since v1)

Field	Type	Description
id	string	Stable unique identifier, prefixed by source.
messages	list	Ordered turns, each {role, content} with role ∈ {user, assistant}. Since v2, may contain more than 2 turns (strict user/assistant alternation) for the multi-turn sources — see below.
source	string	Origin, one of the 12 pool sources — see Dataset Composition.
layer	string	instruct or creatif.
lang	string	Always fr.
needs_lid	bool	Whether language still needed verification during processing.

No system prompt is stored in the data. A system prompt (anchoring the creative, no-refusal persona) is intended to be injected at collation time, so it can be varied and applied uniformly across all sources rather than becoming a proxy for one domain.

Dataset Composition

Pool total (v2): 134,954 examples, 12 sources. By layer: instruct 127,228 (94.3 %), creatif 7,726 (5.7 %). Multi-turn examples (≥2 assistant turns): 1,616 (1.2 %). (Source: project QC metrics on selene.jsonl, 2026-07-24.)

Source	Layer	n	%	Role	Origin
french_alpaca	instruct	109,148	80.9	General French instruction-following	jpacifico/French-Alpaca-dataset-Instruct-110K
extra_openhermes_fr	instruct	7,994	5.9	Diverse instruction tasks (rewriting, summarization, classification, explanation, some code), curated (v2)	legmlai/openhermes-fr
extra_cfp	instruct	6,000	4.4	Native French generative tasks — rewriting, simplification, titling, paraphrase, summarization (19 task types) (v2)	CATIE-AQ/CFP
french_pd_books	creatif	5,435	4.0	Public-domain literary prose, single-turn (Project Gutenberg)	CATIE-AQ/french_books
oasst2	instruct	1,958	1.5	French assistant conversation pairs	OpenAssistant/oasst2
extra_aya	instruct	1,341	1.0	Native-human-written French instructions (v2)	CohereLabs/aya_dataset (fra subset)
wikisource	creatif	1,298	1.0	Public-domain literary prose, single-turn, fiction-filtered	manu/wikisource_fr
french_pd_chain	creatif	824	0.6	Multi-turn "continue" — the follow-up turn is the true, consecutive next segment of the same book (v2)	Derived from Project Gutenberg text (re-segmented from the same source as french_pd_books)
oasst2_thread	instruct	787	0.6	Reconstructed full multi-turn French oasst2 conversation threads, not just isolated pairs (v2)	OpenAssistant/oasst2
french_pd_long	creatif	153	0.1	Long-form, single-turn, generated under verifiable constraints (POV, length, tone, genre, character); responses up to ~2,700 words (v2)	Derived from Project Gutenberg text
french_pd_conte	creatif	11	0.0	Complete public-domain short stories / contes (v2)	Derived from Project Gutenberg text
wikisource_chain	creatif	5	0.0	Multi-turn "continue" — consecutive segments of the same Wikisource work (v2)	Derived from manu/wikisource_fr

Sources marked "(v2)" are new since the first release. french_pd_chain, french_pd_long and french_pd_conte are built from the same underlying Gutenberg corpus as french_pd_books but re-segmented to preserve consecutive book segments and longer spans; passages consumed by a chain are excluded from the single-turn pool to avoid duplication. The pool is dominated by french_alpaca (80.9 %); use the layer field, and the project's build script, to obtain a balanced training mix. Purely synthetic creative data may be added later.

Dataset Creation

Curation Rationale

The project follows a base → instruct from scratch route: no refusal is ever shown to the model, so the resulting model is uncensored-for-fiction by construction rather than by stripping guardrails afterwards. The creative layer does the specialization work, which is why most of the effort goes into sourcing clean literary prose and turning it into faithful instruction pairs.

Source Data

• French-Alpaca — a flat, 100 % French instruction dataset; used as-is for the general instruction layer.
• oasst2 — a multilingual conversation tree. French prompter → assistant pairs (oasst2) were reconstructed via parent links and kept only when detected as French; as of v2, full multi-turn French threads (oasst2_thread) are also reconstructed from the same tree (see below).
• CATIE-AQ/french_books — French public-domain books from Project Gutenberg, human-proofread (Distributed Proofreaders), used for the literary creative layer (french_pd_books, and, as of v2, re-segmented into french_pd_chain, french_pd_long and french_pd_conte).
• manu/wikisource_fr — text extracted from French Wikisource (community-proofread transcriptions of public-domain works), used as an additional literary source after fiction filtering (wikisource, and, as of v2, wikisource_chain).
• legmlai/openhermes-fr (v2) — OpenHermes (a diverse English instruction corpus, distinct from the Alpaca lineage) translated and then regenerated in French by GPT-4o, with automatic quality flags (bad_entry) provided by the source dataset.
• CATIE-AQ/CFP (v2) — natively French instruction prompts (not translated), template-generated over French digital-native corpora (e.g. Alloprof, CLIPS/MFAQ, Wikipedia:Oracle, FrenchMedMCQA), covering 37 task types in the source.
• CohereLabs/aya_dataset (fra subset) (v2) — natively human-written and human-annotated French instructions from Cohere For AI's Aya platform (not a translation, not LLM-generated, not OCR).

Instruction Backtranslation

The single-turn creative pairs (french_pd_books, wikisource, and, as of v2, french_pd_long) are built by instruction backtranslation (Li et al., 2023): because public-domain passages are responses without prompts, we generate the missing instruction from each passage rather than the reverse. A local Ministral-8B-Instruct-2410 — the same model family as the target base model, run in 4-bit on a single 8 GB GPU — reads each excerpt and writes one plausible French writing brief under strict rules: start from a rotated (and genre-aware) opening verb, state the genre / tone / situation, and use no proper noun and no wording taken from the excerpt. The excerpt becomes the assistant response and the generated brief the user instruction. Briefs that copy the passage (n-gram overlap), reuse its names, or are malformed are filtered out afterwards.

v2 constrained-generation variant (french_pd_long) — for long-form passages (responses up to ~2,700 words), the backtranslation prompt was extended to state verifiable constraints as facts detected from the passage itself: point of view, target length, tone, genre, era and main character, so the instruction can later be checked against the delivered text (e.g. "≈2,000 words, first person, anxious tone"). An internal QC pass measured length adherence at 986/986 within ±25 % of the requested word count. Point-of-view adherence was noisier on first release (the requested POV is not always what the underlying 19th-century excerpt actually delivers); 55 instructions were rewritten from the real text to fix clear mismatches, and ~100 ambiguous cases (dialogue- or quote-heavy passages where POV is hard to call) were left as-is — treat POV-following as a softer signal than length-following in this layer.

Multi-Turn Reconstruction (v2)

Three sources depart from the single user/assistant pair format:

• oasst2_thread — instead of flattening the oasst2 conversation tree into isolated prompter → assistant pairs, full French conversation threads are reconstructed by walking parent_id chains end to end and kept only when the thread is in French.
• french_pd_chain and wikisource_chain — consecutive passages from the same public-domain book / work are joined into 2–3-turn conversations: an initial instruction-backtranslated turn, followed by one or two "continue the story" turns (varied phrasings) whose assistant response is the true, consecutive next segment of the same source text (verified programmatically: the follow-up segment matches the next segment index of the same book_id). Segments consumed by a chain are excluded from the corresponding single-turn source to avoid duplicating the same text twice in the pool. Wikisource yielded few such chains (5) because its identifiers key on physical page, not work position, and the sampled pool only sparsely covers full works; Gutenberg's book_id ordering allowed many more (824).

New Instruct Sources (v2)

Three additional French instruction sources were evaluated and added to diversify task types beyond French-Alpaca's short, single-style prompts (full evaluation methodology, per-source scoring and rejected candidates in the project's internal QC notes):

• extra_openhermes_fr — sampled from legmlai/openhermes-fr, with bad_entry-flagged rows dropped, "You are an assistant…" (Orca/FLAN-style) framing prompts removed to avoid reintroducing assistant tics, and a cap on code-related tasks so they do not dominate a creative-writing-oriented instruct layer.
• extra_cfp — sampled from CATIE-AQ/CFP, restricted to a whitelist of generative task types (rewriting, simplification, titling, paraphrase, summarization, …), excluding pure classification tasks (NER/POS/sentiment/entailment) whose outputs are bare labels with little value for a creative model, and round-robin balanced across task types.
• extra_aya — the French (fra) subset of CohereLabs/aya_dataset, used close to as-is given its small size and native-human origin.

All three were checked for overlap against a French-Alpaca sample (exact + MinHash); measured overlap was near zero for all three sources before the pool-wide deduplication described below was applied on top.

Data Collection and Processing (methods)

All sources are normalized to a common ChatML-style schema. During training, loss is expected to be masked on the prompt/system turns and computed only on the assistant turn(s); as of v2 some conversations have more than one assistant turn (see Dataset Structure).

General instruction layer

• Removal of refusals and moralizing answers (assistant responses matching refusal / "as an AI" / "not appropriate" patterns are dropped) — enforcing the zero-refusal design.
• Language filtering with a fastText language-identification model applied to the (longer, more reliable) assistant response; a confidence threshold controls strictness.
• For conversation trees, pairs and full threads are reconstructed from parent links before filtering.

Creative layer — Project Gutenberg

• Sampling of fiction authors from a curated allow-list, with a per-author book cap to prevent any single prolific author from dominating the literary voice.
• Cleaning: removal of transcription notes, title pages, tables of contents and colophons via end-anchor detection and a dynamic head trim; stripping of inline page numbers, note markers and italic markup.
• Segmentation on chapter boundaries (chapter/part headings and section numerals are consumed as separators), then sentence-aware grouping into passages of roughly 200–900 words, with abbreviation-protected sentence splitting.
• Quality filters at passage level (bounds on uppercase, digit and "garbage" token ratios) plus a language check.
• Instruction backtranslation: a local instruction model generates one plausible writing brief per passage. The generation prompt forbids copying any wording, forbids proper nouns from the excerpt, and rotates the opening verb/register to diversify brief form. Passages remain the assistant responses; the generated briefs become the instructions.
• Post-generation curation: n-gram overlap filter (reject briefs that leak the passage), proper-noun-leak filter, and removal/repair of malformed briefs.
• A final per-author passage cap balances the literary voice across authors.

Creative layer — French Wikisource

• Metadata pre-filter: identifiers encode the work title/author, so periodicals and reference works (reviews, journals, dictionaries, …) are discarded before any model call — cheaper and more precise than classifying them.
• Wikitext cleaning that preserves paragraph structure (unlike the flattened Gutenberg source), giving properly paragraphed responses.
• Fiction classification of each candidate passage using the local model with structured output / constrained decoding: the model is forced to emit a schema-validated object (fiction boolean + literary genre), and only genuine narrative fiction with a consistent literary genre is kept.
• Genre-aware instruction backtranslation (verse and drama get dedicated brief forms), followed by the same curation as above.

Reinforced pool-wide curation (v2)

• Refusal / assistant-tic purge: an expanded refusal regex flagged residual matches inherited from the general instruction sources; after manual inspection, 115 were confirmed as genuine refusals or assistant tics anchored at the start of the response and removed (52 french_alpaca, 29 oasst2, 27 oasst2_thread, 6 extra_openhermes_fr, 1 french_pd_books). Fiction-dialogue lines such as a character saying "I can't" are not refusals and were kept, per the project's zero-refusal rule (a refusal is the assistant declining the task, not fictional speech).
• Editing-artifact removal: a filter targeting leftover tables of contents, "Agrandir" image captions and similar transcription residue removed 308 examples from the pool.
• POV-fidelity correction: see the french_pd_long note above (55 instructions rewritten from the source text).
• Net effect: the merge produced 135,377 raw lines from 12 sources; after the refusal purge (−115) and artifact removal (−308) the published pool stands at 134,954.

Merged pool → training set (build step)

• A deterministic, idempotent build turns the merged pool into a training set: a response-length filter, exact + MinHash near-duplicate deduplication, a creative response-level dedup (one passage kept per unique target), a v2 cross-format dedup on assistant content applied before the split (catches the same underlying text reappearing under different framings — e.g. an oasst2 pair and an oasst2_thread turn sharing an answer, or a wikisource passage reused inside a wikisource_chain — which had produced measurable train/validation leakage before this step was added), ratio balancing (keep all creative, subsample instruct via the layer field to the target share), and a 90/10 train/validation split. Decontamination against evaluation prompts is applied before training. The published file is the raw merged pool; the balanced split is produced from it and is not itself published — the current build produces roughly 14.8k train / 1.6k validation examples at a 45 % creative share, per the project's own build logs.

Annotations

The instructions in the creative layer are machine-generated (instruction backtranslation with a local instruction model), including the "continue" turns of the multi-turn chains (varied phrasings, not model-generated per se, but templated by the pipeline). The responses are human-authored public-domain literary text in every creative source. Fiction/non-fiction labels on the Wikisource layer are also model-generated (schema-constrained). In the instruction layer, extra_aya and CFP instructions are human/template-authored in their source datasets, while extra_openhermes_fr instructions and responses are LLM-regenerated (GPT-4o) in the source dataset.

Personal and Sensitive Information

The creative layer is public-domain literature. The instruction layers are public, widely used datasets. No deliberate collection of personal data was performed; standard risks of large text corpora still apply.

Bias, Risks, and Limitations

• Uncensored for fiction. The dataset intentionally contains no refusals and includes mature or dark fictional themes handled straight, within a fiction frame. A model trained on it will follow such fiction instructions. Downstream users are responsible for deployment safeguards and for complying with applicable law; the dataset is not intended to produce real-world harmful, illegal, or non-consensual content about real people.
• Dated register. The literary layer is drawn from 19th–early-20th-century public-domain prose: a rich literary voice but an archaic register, to be balanced with more modern creative data.
• Machine-generated instructions. Backtranslated briefs can occasionally drift from the passage or read as templated; curation reduces but does not eliminate this.
• Raw, unbalanced pool. The published file is instruction-heavy (94.3 % instruct); use the layer field, and the project's build script, to obtain the intended ~45 % creative training ratio.
• POV adherence is a softer signal than length adherence (v2, french_pd_long). Because responses are real literary excerpts rather than generated text, a requested point of view is not always fully honored by the underlying passage. 55 clearly-wrong instructions were corrected against the source text, but ~100 ambiguous cases were left as-is; do not treat first-/third-person compliance in this layer as a clean, fully verified signal the way response length is (986/986 within ±25 % per internal QC).
• Multi-turn examples remain a small minority of the pool (about 1.2 % of examples have ≥2 assistant turns), concentrated in oasst2_thread, french_pd_chain and wikisource_chain. They are a real but limited signal for training multi-turn coherence and "continue the story" behavior.
• Large-scale exact/MinHash and cross-format deduplication happen at training-build time, not on the raw pool as published; users building their own training set from this pool should apply equivalent deduplication before splitting into train/validation, particularly across sources that share underlying text (e.g. oasst2/oasst2_thread, wikisource/wikisource_chain, french_pd_books/french_pd_chain).

Uses

Direct Use

Supervised fine-tuning of French creative-writing / storytelling / role-play assistants, for research and non-commercial purposes. The intended target is a from-scratch instruct model trained on the Ministral-3-14B base (Mistral family), later quantized aggressively (IQ2/IQ3) to fit an 8 GB consumer GPU.

Out-of-Scope Use

Any commercial use (see licensing), and any use aimed at generating real-world harmful, illegal, or non-consensual content, or content targeting real individuals.

Licensing Information

Selene is a derivative of multiple sources and a model-generated component. The effective terms are the union of the most restrictive: non-commercial, research-only, with attribution and share-alike obligations.

Component	License / status
French-Alpaca (french_alpaca)	Apache-2.0
oasst2 (oasst2, oasst2_thread)	Apache-2.0
Project Gutenberg texts (via CATIE-AQ/french_books; french_pd_books, french_pd_chain, french_pd_long, french_pd_conte)	Public domain (US); source dataset declares no explicit license
French Wikisource text (via manu/wikisource_fr; wikisource, wikisource_chain)	Underlying works public domain; Wikimedia transcriptions under CC BY-SA; source dataset declares no explicit license
openhermes-fr (extra_openhermes_fr, v2)	ODC-BY 1.0
CATIE-AQ/CFP (extra_cfp, v2)	Aggregate of its constituent sub-corpora's licenses, predominantly permissive (CC-BY-family); no single blanket license declared — see the source dataset card for the per-sub-corpus breakdown
Aya Dataset, fra subset (extra_aya, v2)	Apache-2.0
Instruction briefs & fiction labels	Generated with Ministral-8B-Instruct-2410 under the Mistral Research License (research, non-commercial)

Because the creative-layer instructions were produced with a model under the Mistral Research License, the dataset as a whole must be treated as research / non-commercial, regardless of the more permissive licenses (Apache-2.0, ODC-BY) covering some of the v2 instruction sources. Attribution to all sources is required; the CC BY-SA share-alike condition applies to the Wikisource-derived portion, and the ODC-BY attribution condition applies to the openhermes-fr-derived portion. If you reuse Selene, comply with all upstream licenses listed above, including the CFP sub-corpus licenses.

Citation

If you use Selene, please cite this dataset and the upstream sources.

@misc{selene_dataset,
  title  = {Selene: A French Creative-Writing Instruction Dataset},
  author = {Chortan},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Chortan/selene}}
}

References

• French-Alpaca — jpacifico. https://huggingface.co/datasets/jpacifico/French-Alpaca-dataset-Instruct-110K
• OpenAssistant Conversations (oasst2) — Köpf et al., OpenAssistant Conversations — Democratizing Large Language Model Alignment, NeurIPS 2023. https://huggingface.co/datasets/OpenAssistant/oasst2
• French public-domain books — CATIE-AQ / Project Gutenberg. https://huggingface.co/datasets/CATIE-AQ/french_books · https://www.gutenberg.org
• French Wikisource — https://fr.wikisource.org · https://huggingface.co/datasets/manu/wikisource_fr
• openhermes-fr — legmlai. https://huggingface.co/datasets/legmlai/openhermes-fr
• CATIE-AQ/CFP — CATIE. https://huggingface.co/datasets/CATIE-AQ/CFP
• Aya Dataset — Singh et al., Aya Dataset: An Open-Access Collection for Multilingual Instruction Tuning, 2024. https://huggingface.co/datasets/CohereLabs/aya_dataset
• Ministral-8B-Instruct-2410 — Mistral AI. https://huggingface.co/mistralai/Ministral-8B-Instruct-2410
• Instruction backtranslation — Li et al., Self-Alignment with Instruction Backtranslation, 2023.
• Self-Instruct — Wang et al., Self-Instruct: Aligning Language Models with Self-Generated Instructions, 2022.
• Language identification — Joulin et al., fastText (lid.176), 2016.

Changelog

v2 — 2026-07-24

• Pool grown from 118,121 to 134,954 examples, from 4 to 12 sources.
• New multi-turn layers: oasst2_thread (787, reconstructed full French oasst2 threads), french_pd_chain (824, true consecutive Gutenberg "continue" chains), wikisource_chain (5, same for Wikisource). Multi-turn examples now make up 1.2 % of the pool (0 in v1, which was 100 % single-turn).
• New constrained long-form single-turn layer: french_pd_long (153 examples, responses up to ~2,700 words, instructions encoding verifiable POV / length / tone / genre / character constraints via a v2 backtranslation prompt).
• New complete-short-story layer: french_pd_conte (11 examples).
• Three new instruction sources added to diversify task types beyond French-Alpaca: extra_openhermes_fr (7,994), extra_cfp (6,000), extra_aya (1,341); a fourth candidate (evol-instruct-french) was evaluated and rejected (Alpaca-lineage risk, residual English responses, unclear licensing).
• Curation: 115 residual refusals/assistant tics purged (anchored at the start of the response; fiction-dialogue "I can't" lines kept), 308 editing-artifact examples removed (tables of contents, transcription residue), 55 creative instructions rewritten to fix point-of-view mismatches against the actual source text.
• Schema unchanged (id, messages, source, layer, lang, needs_lid), but messages can now contain more than two turns for the multi-turn sources — see Dataset Structure.
• Licensing: three components added — ODC-BY 1.0 (extra_openhermes_fr), Apache-2.0 (extra_aya), and an aggregate of predominantly permissive sub-corpus licenses (extra_cfp). Overall effective terms are unchanged (still governed by the Mistral Research License via the model-generated instruction/label components).
• The recommended training build (prepare_dataset.py, 45 % creative target) also gained a cross-format response deduplication step applied before the train/validation split, to prevent the same underlying text (e.g. shared between oasst2 and oasst2_thread, or wikisource and wikisource_chain) from leaking across the split.

Acknowledgements

Built on the work of the Project Gutenberg and Wikisource proofreading communities, OpenAssistant contributors, the Cohere For AI Aya community, the CATIE-AQ and openhermes-fr maintainers, and the maintainers of the upstream datasets and models cited above.