formosa-nlu-synth-v1
收藏资源简介:
FormosaNLU Synth 是一份以正體中文(台灣,zh-TW)為主的口語自然語言理解(NLU)合成訓練資料集,涵蓋 60 種意圖(intent)與 55 種槽位類型(slot type)。資料由本機 open-weight 教師模型(Qwen3.6:27b)生成,經 deterministic F1–F6 過濾器與不同家族的獨立評審(F7)稽核後,發布 3,754 筆訓練樣本。每筆資料包含 id、utterance(utt)、intent 標籤、槽位跨度(slots)、風格(style,如 massive_like 或 tw_colloquial)、生成配方(recipe)、教師模型、提示版本、種子樣本 ID、生成參數與過濾分數等欄位。資料來源為 Amazon Science MASSIVE 的 zh-TW 子集,經每意圖約 20 筆種子取樣(共 1,176 筆)後,由教師模型生成 11,264 筆原始數據,經層層過濾最終保留 3,754 筆。下游評估顯示,在完全未參與生成或訓練的 MASSIVE zh-TW 測試集(2,974 筆)上,使用 Gemma 4 E4B QLoRA 模型,過濾後的合成資料相較僅使用真實資料在意圖準確率上平均提升 +4.14 個百分點(95% CI [+2.60, +5.59]),精確匹配提升 +3.86 個百分點(95% CI [+2.75, +4.92]),並在 Phi-4-mini-instruct 模型上複製出類似結果。資料集適合用於正體中文意圖分類與槽位填充研究、合成資料過濾與蒸餾、低資源增強、NLU 管線與穩健性實驗。限制包括:種子資料源自翻譯型 MASSIVE,無法代表所有自然台灣口語;合成記錄可能保留教師偏差或未偵測的品質問題;穩健性測試為確定性探針,非真實 ASR 輸出;不應用於醫療、法律、金融或安全決策。資料集採用 CC BY 4.0 授權。
FormosaNLU Synth is a synthetic training dataset for spoken Natural Language Understanding (NLU) primarily in Traditional Chinese (Taiwan, zh-TW), covering 60 intents and 55 slot types. The data is generated by a local open-weight teacher model (Qwen3.6:27b), filtered through deterministic F1–F6 filters and audited by independent reviewers from different families (F7), resulting in 3,754 training samples. Each sample includes fields such as id, utterance (utt), intent label, slot spans, style (e.g., massive_like or tw_colloquial), generation recipe, teacher model, prompt version, seed sample ID, generation parameters, and filtering scores. The data source is the zh-TW subset of Amazon Science MASSIVE, with approximately 20 seed samples per intent (1,176 total) used to generate 11,264 raw samples from the teacher model, which are then progressively filtered to retain 3,754 samples. Downstream evaluation on the MASSIVE zh-TW test set (2,974 samples), which was not involved in generation or training, shows that using Gemma 4 E4B QLoRA with the filtered synthetic data improves intent accuracy by an average of +4.14 percentage points (95% CI [+2.60, +5.59]) and exact match by +3.86 percentage points (95% CI [+2.75, +4.92]) compared to using only real data, and similar results were replicated on the Phi-4-mini-instruct model. The dataset is suitable for research on Traditional Chinese intent classification and slot filling, synthetic data filtering and distillation, low-resource augmentation, NLU pipelines, and robustness experiments. Limitations include: seed data originates from translated MASSIVE, which cannot represent all natural Taiwanese spoken language; synthetic records may retain teacher biases or undetected quality issues; robustness tests are deterministic probes, not real ASR outputs; and it should not be used for medical, legal, financial, or safety decisions. The dataset is licensed under CC BY 4.0.
FormosaNLU Synth 数据集概述
基本信息
- 语言:正體中文(台灣,
zh-TW) - 许可证:CC BY 4.0
- 任务类别:文本分类、词元分类(意图识别、槽位填充)
- 数据规模:1K < n < 10K,共 3,754 笔训练数据
- 标签体系:涵盖 60 种 intent 与 55 种 slot type
- 数据风格:包含
massive_like与tw_colloquial两种语料风格
数据内容与结构
- 文件组成:
data/train.jsonl:3,754 行训练数据schema.json:JSON Schema 定义release_manifest.json:来源 artifact、SHA-256、笔数与版本信息
- 每条记录字段:
id:稳定 synthetic sample IDutt:正體中文(台灣)utteranceintent:MASSIVE intent labelslots:{type, value}slot spansstyle:massive_like或tw_colloquialrecipe:synthetic recipeteacher_model:生成模型prompt_version:prompt recipe versionseed_sample_id:来源 MASSIVE seed identifiergeneration_params:frozen generation parametersfilter_scores:F5/F6 contamination 与 similarity evidence
数据来源与生成流程
- 上游种子数据:Amazon Science MASSIVE
zh-TW,每 intent 约 20-shot,共 1,176 笔 - 生成模型:
qwen3.6:27b,本地 Ollama、open-weight - 生成规模:
- 原始生成:11,264 笔
- F1–F6 过滤后:3,760 笔
- F7 发布:3,754 笔
- 过滤机制:包含 schema、label/slot span、script/language、duplicate、seed-copy、Test contamination、embedding outlier 与 independent judge audit;F7 random stratum 观察漏检率为 6.0%(样本仅 50 笔)
下游评估证据
第一轮(Gemma 4 E4B QLoRA)
在 2,974 行的 MASSIVE zh-TW Test 上,三个 paired seeds(42–44)显示:
- Intent accuracy:filtered 相比 real-only 平均提升 +4.14 ± 1.39 个百分点
- Exact match:平均提升 +3.86 ± 0.73 个百分点
- Robustness:seed-42 filtered adapter 在 ASR-noise、colloquial、lexical 三种 probes 上的 intent accuracy、slot F1、exact match 均高于 real-only
- 统计检验:5,000 次 hierarchical paired bootstrap 后,intent accuracy 95% CI [+2.60, +5.59],exact match 95% CI [+2.75, +4.92];每个 seed 的 McNemar tests 经 Holm correction 后均
p ≤ 0.00017
第二轮(Phi-4-mini-instruct 复制)
| Metric | Gemma Δ [95% CI] | Phi Δ [95% CI] |
|---|---|---|
| Intent accuracy | +4.14 [+2.60, +5.59] | +5.09 [+1.83, +9.02] |
| Exact match | +3.86 [+2.75, +4.92] | +4.71 [+1.36, +7.59] |
判准(两个 family 均需正向平均提升且 CI 下界大于零)通过,显示效益不是单一 student model 的特性。
Robustness probe 复制
在 8,922 笔扰动 probe 上,两个 family 各三个 seed 的 paired delta(百分点):
| Metric | Gemma Δ ± SD | Phi Δ ± SD |
|---|---|---|
| Intent accuracy | +3.63 ± 1.72 | +6.22 ± 3.46 |
| Exact match | +3.58 ± 2.05 | +6.98 ± 3.29 |
十项全部为正,但强度不同;Phi 的五项 mean 均大于各自 SD,Gemma 仅两项如此。作者声明十个正数不等于十个已确立的效果。
未达标的指标
json_valid_rate未通过预设门槛(Gemma 侧 CI 跨越零)- Phi 的
exact_match_seed_42在 Holm 校正后p = 0.141不显著
适用场景
- 正體中文 intent classification/slot filling 研究
- Synthetic-data filtering、distillation 与 low-resource augmentation
- NLU pipeline、schema-constrained output 与 robustness 实验
限制与不适用场景
zh-TWseed 源自翻译型 MASSIVE,无法代表所有自然台湾口语- Synthetic records 可能保留 teacher bias 或未偵测品质问题
- F7 随机层漏检率为 6.0%,但样本仅 50 笔,置信区间很宽
- Robustness 为 deterministic probes,非真实 ASR log
- 不应将 intent/slot predictions 用于医疗、法律、金融或安全决策
- 未发布 MASSIVE Test/validation,使用者需自行依 MASSIVE 授权取得评估资料
授权与引用要求
- 数据集使用 CC BY 4.0,上游 MASSIVE
zh-TW同样为 CC BY 4.0 - 使用或再发布时需:标柱 MASSIVE 与 FormosaNLU Synth;链接 CC BY 4.0;说明经 synthetic generation 与 filtering 修改
- 引用格式:
- MASSIVE:Jack FitzGerald et al. ACL 2023
- FormosaNLU Synth:kuotunyu. "FormosaNLU Synthetic Data Distillation for Traditional Chinese (Taiwan) NLU, version 1.0.0. 2026."(完整代码、决策记录与实验报告见 GitHub 仓库)
可重现性说明
release_manifest.json 提供发布档 SHA-256、原始 release artifact SHA-256、Git source commit 与确切笔数;完整生成、过滤与评估程序位于 GitHub 仓库中。





