asd567557275/zhtw-roleplay-space-grimoire

Space Grimoire RP Corpus (Traditional Chinese)

Speaker-attributed dialogue from the original novel 空間魔導書與少年魔法師 (The Space Grimoire and the Young Mage; 283 chapters, ~1.7M characters), cut into scenes and assembled into ShareGPT-style role-play training data. The novel and this dataset are the work of 睡半夜怎麼三更, who holds the copyright and has no exclusive platform agreement. Data: CC BY 4.0. Code: Apache 2.0.

中文說明在下方

Dataset Summary

Supported Tasks

Character role-play fine-tuning (multi-turn, one target character per sample), speaker attribution, dialogue tone analysis.

Languages

Traditional Chinese (zh-Hant, Taiwan usage).

Dataset Structure

Data Instances

ShareGPT row (sft/*_sharegpt.jsonl):

{"conversations": [
   {"from": "system", "value": "你是《空間魔導書與少年魔法師》裡的芙雷德莉卡。現在的時間點是序卷・序章 蒸氣列車，只知道到這裡為止發生的事。……"},
   {"from": "human",  "value": "（場景：深夜，蒸氣列車商務車廂）\n維吉尼亞：「哎呀……！」\n……"},
   {"from": "gpt",    "value": "*芙雷德理所當然的回答。* 「我拒絕，空間是非常危險的力量……」"}],
 "character": "芙雷德莉卡", "chapter": 1, "part": "序卷", "scene_id": "sc00000", "n_gpt_turns": 4, "split": "train"}

Data Fields

data/utterances.jsonl, one row per quoted segment:

data/scenes.jsonl: scene_id, chapter, part, pid_start, pid_end, break_reason, setting_hint, participants, uids.

data/paragraphs.jsonl: the full novel text, one paragraph per row: pid, part, chapter, chapter_in_part, chapter_title, role (body, chapter_title, part_title, front), text.

sft/flat_utterances.jsonl: one target-character line per row with the previous five turns as context, intervening narration, setting_hint, tone, addressee.

Data Splits

Scenes are split 9:1; train and validation never share a scene.

Characters in the SFT files

SFT filter: speaker_type == named, modality in {spoken, broadcast}, and not marked uncertain by review. Edit targets in src/p5_config.json and rerun python src/p5_sft.py to change the character set.

Dataset Creation

Source Data

The author's final manuscript (.docx, 2026-09-12 revision), converted to one paragraph per line; the front-matter contact line was removed. Paragraph ids (pid) are stable across all files in this repository.

Annotations

1. 1.Labelling: Claude Opus 5 labelled every chapter with the schema above, citing an 8–15 character evidence span per line.
2. 2.Review: Claude Fable 5.1 reviewed a sample of 1,721 lines (all unresolved and flagged lines plus 10% of the rest): 1,574 ok, 101 wrong, 46 uncertain. Most "wrong" verdicts were __unknown__/__anon__ labels that the reviewer resolved to a name. Only 15 lines (0.9%) carried a wrong name; all were characters labelled as bystanders before their name is revealed, and all are corrected.
3. 3.Author rulings: the author personally ruled 52 lines; 15 more were filled from unambiguous context. Rulings are listed in data/adjudications.json and applied by src/p6_adjudicate.py.

Pipeline code and prompts are in src/.

Personal and Sensitive Information

None. The text is fiction.

Considerations for Using the Data

• —Characters who appear before their name is revealed may occasionally still be labelled __anon__ (estimated < 1%). alias_used supports cross-chapter identity linking.
• —Chapter 247 did not go through the sampled review.
• —tone is a model judgement, not ground truth.
• —The SFT system prompt frames each sample at its point in the story ("you only know what has happened so far"); characters' knowledge and stance change across the nine parts.

Additional Information

• —Author: 睡半夜怎麼三更
• —License: data CC BY 4.0, code Apache 2.0
• —Training example: train/train_unsloth.py (Qwen3-8B 4-bit QLoRA, loss on gpt turns only), see train/README.md.

Citation

@misc{space_grimoire_rp_2026,
  title  = {Space Grimoire RP Corpus (Traditional Chinese)},
  author = {睡半夜怎麼三更},
  year   = {2026},
  note   = {Speaker-attributed dialogue and ShareGPT role-play data from the novel 空間魔導書與少年魔法師},
  license = {CC-BY-4.0}
}

《空間魔導書與少年魔法師》繁體中文角色扮演語料

原創長篇小說《空間魔導書與少年魔法師》（283 章、約 174 萬字）全書對白的說話者標註，切成場景後組成 ShareGPT 格式的角色扮演訓練資料。小說與本資料集皆為 睡半夜怎麼三更 的作品，版權作者自有，未與任何平台簽訂獨家授權。資料採 CC BY 4.0，程式採 Apache 2.0。

概要

用途

角色扮演微調（多輪、一列一個目標角色）、說話者歸屬、對白語氣分析。

資料結構

範例

ShareGPT 列（sft/*_sharegpt.jsonl）：system 帶角色名、篇章時間點（「只知道到這裡為止發生的事」）與假名提示；human 是場景描述、敘述與其他人的台詞；gpt 是目標角色的台詞，前面用 *…* 帶原文的動作描寫。超過 40 句的場景切成 30 句一窗、重疊 5 句。範例見上方英文段。

欄位

data/utterances.jsonl，一列一個引號片段：

data/scenes.jsonl：scene_id、chapter、part、pid_start、pid_end、break_reason、setting_hint、participants、uids。

data/paragraphs.jsonl：小說全文，一段一列：pid、part、chapter、chapter_in_part、chapter_title、role（body / chaptertitle / parttitle / front）、text。

sft/flat_utterances.jsonl：一句一列，帶前五句 context、中間敘述、setting_hint、tone、addressee。

切分

依場景 9:1，train 與 validation 不共用場景。

SFT 收錄角色

威爾斯 4,064 句、芙雷德莉卡 2,386、質麗公主 1,042（多數章節用假名拉洛莉亞）、禮西奈 795、珊德菈 516、奧蘿菈 427、霄 260、米哈伊爾 172、海因里希 170、拉娃 157、夏綠蒂 145、蓮華 118。過濾規則：speaker_type=named、modality 為 spoken 或 broadcast、審核未判不確定。改 src/p5_config.json 的 targets 後重跑 python src/p5_sft.py 即可換角色。

製作流程

原文

作者定稿 .docx（2026-09-12 修訂版）轉成一段一行，前言的聯絡資訊已移除。段落編號 pid 在本資料集所有檔案間一致。

標註

1. 1.標註：Claude Opus 5 逐章標註，每句引用 8–15 字原文依據。
2. 2.審核：Claude Fable 5.1 抽審 1,721 句（未定句、旗標句必審，另抽一成）：ok 1,574、wrong 101、uncertain 46。wrong 多數是 __unknown__／__anon__ 被審核補回正名；「標了名字但錯」15 句（0.9%），全部是角色報名字前被標成路人，已修正。
3. 3.作者裁定：作者親自裁定 52 句，另 15 句依明確前後文回填。清單在 data/adjudications.json，由 src/p6_adjudicate.py 套用。

管線程式與 prompt 在 src/。

個人資訊

無，內容為虛構小說。

使用注意

• —報名字前出場的角色可能仍有少數被標成 __anon__（估計 <1%），可用 alias_used 做跨章連結。
• —第 247 章未經抽樣審核。
• —tone 是模型判斷，不是真值。
• —system prompt 把每個樣本定在故事的當下時間點，角色的認知與立場在九篇之間會變。

其他

• —作者：睡半夜怎麼三更
• —授權：資料 CC BY 4.0，程式 Apache 2.0
• —訓練範例：train/train_unsloth.py（Qwen3-8B 4bit QLoRA，只對 gpt 輪算 loss），說明見 train/README.md
• —引用格式見上方英文段。