Update data card to wvs-aligned format

README.md

@@ -1,325 +1,175 @@
-个人博客中非结构化数据提取数据。
-  每个data中包含大约800个个体。
-  每个个体包含的数据量大小从20-2000不等，大部分个体observation为50-150之间。
-  大约有5%的个体observation数量>800.大样本意味着对个体的更具体观察，值得探索大样本的更多利用空间。
-  每个问题包含四个不同选项。
-  每个个体的observation：question==3:1
+---
+language:
+  - en
+license: other
+pretty_name: Chiron Blog Authorship
+tags:
+  - blog
+  - personality
+  - psychology
+  - mcq
+size_categories:
+  - 1K<n<10K
+configs:
+  - config_name: default
+    data_files:
+      - split: train
+        path: qa/train-*.parquet
+      - split: val
+        path: qa/val-*.parquet
+      - split: test
+        path: qa/test-*.parquet
+---
 
-数据来源： Blog Authorship Corpus   https://tianchi.aliyun.com/dataset/90243
-该数据集整理了一段时间内不同个体的博客内容，愿数据集总计68万条博客。
-原始数据形式： 【id,gender,age,topic,sign,date,text】 (同一ID表示同一个人)
-例如：第一条数据 2059027,male,15,Student,Leo,"14,May,2004"，"   Info has been found (+/- 100 pages, and 4.5 MB of .pdf files) Now i have to wait untill our team leader has processed it and learns html.   "
+# Chiron-Blog
 
-数据处理使用模型  deepseek-v3.2   随机数种子：106   
-考虑到该数据集非结构化，个人色彩较强，在总结人格特质和生成问题时需要模型一定的发散性思维。observation和question的处理中设置temperature为0.6  
-抽样多个个体进行多次提取，observation数量差别<=3%.各observation和question之间语义基本一致。
+基于 **Blog Authorship Corpus** 个体级博客文本整理而成的数据集,服务于 **Chiron** 测量-预测流水线 —— 从受访者已知响应中合成心理学测量,再据测量预测留出响应。
 
-处理流程：
-1，由于总数量过大，将总数据集按每20000条切分处理
+本数据卡是**自包含**的:Parquet 文件、字段 schema、生成策略、上游来源与许可,以及完整复现 / 反向转换脚本都直接随仓库附带。
 
-2，按照ID号分为不同个体，提取出gender，age，topic（工作类型，仅部分人包含）特征。忽略sign（星座），date（博客发布日期）。
-  考虑到单个个体需要足够数量的信息，个体博客数量小于20条的跳过，不再进行后续处理。
+## 文件清单
 
-3，个体之间串行，对每个个体的每条博客并行分析，提取出与人格特质相关的描述，称为观测observation。每条博客中可提取多条观测。若无法体现出人格特质则不提取。
-  保留该过程得到的中间结果
+```
+data_01.zst ... data_07.zst    上游 zstd 压缩 JSON 数组,每个元素是一名个体
+                                 (observations + predictions + extra_info)
+                                 共 5 858 名个体
+qa/
+  train-00000-of-00001.parquet   4 366 行
+  val-00000-of-00001.parquet       243 行
+  test-00000-of-00001.parquet      243 行
+scripts/
+  build_qa.py             data_*.zst -> qa/*.parquet
+  parquet_to_jsonl.py     任意 parquet -> jsonl(自动处理 ZSTD/Snappy/GZIP)
+```
 
-4，对每个个体，得到的观测若<=15,数量过小直接舍弃。对其余个体的观测用random=106随机打乱后选靠后的1/4observation提取question。
-  对每个观测，将其重点反映的个人特质转变成选项式的问题。选项之间保持较为细粒度的差异，或者在其反映出的背后特质上有差异。
-  在得到回答后用随机数的方式打乱四个选项，保证随机性。
+所有 Parquet 均使用 ZSTD 压缩 + 列式存储 + 字典编码。
 
+## 加载方式
 
-3，4 部分的具体prompt如下： 
+```python
+from datasets import load_dataset
 
-3，PROMPT_TEMPLATE
-"""You are a research assistant trained in computational social science.
+ds = load_dataset("chiron-group/Blog_Authorship_Corpus")
+print(ds)              # train / val / test
+print(ds["train"][0])
+```
 
-Your task is to extract ALL HIGH-CONFIDENCE personality-related features from a single blog post.
+或直接用 `pandas` / `pyarrow`:
 
----
+```python
+import pandas as pd
+train = pd.read_parquet("hf://datasets/chiron-group/Blog_Authorship_Corpus/qa/train-00000-of-00001.parquet")
+val   = pd.read_parquet("hf://datasets/chiron-group/Blog_Authorship_Corpus/qa/val-00000-of-00001.parquet")
+test  = pd.read_parquet("hf://datasets/chiron-group/Blog_Authorship_Corpus/qa/test-00000-of-00001.parquet")
+```
 
-STRICT RULES:
+## QA Schema
 
-1. You must base your conclusions ONLY on explicit textual evidence.
-2. You must NOT speculate or infer beyond what is clearly supported.
-3. You may extract MULTIPLE features, but ONLY if each one is strongly supported.
-4. If no feature is strongly supported, return an empty list.
-5. Avoid clinical or medical diagnosis.
-6. Be conservative: when in doubt, do NOT include the feature.
+切分粒度是**个体级**:同一个体的所���题目都落在同一个 split 中。
 
----
+### `train`
 
-DEFINITION:
+| 字段 | 类型 | 含义 |
+|---|---|---|
+| `observations` | `list[str]` | 该个体的第一人称行为 / 兴趣 / 倾向陈述,作为测量模型的输入。 |
+| `predictions` | `list[{question, ground_truth}]` | 该个体 ~2/3 的 MCQ 题项。用于测量质量打分以及 prediction-SFT 的监督。 |
+| `held_out_predictions` | `list[{question, ground_truth}]` | 剩余 ~1/3。**训练阶段绝不使用**,留作 within-respondent 留出评估。 |
+| `data_source` | `str` | 固定为 `"Blog Authorship Corpus"`。 |
+| `split` | `str` | `"train"`。 |
+| `extra_info` | `struct` | `{interview_id: str, n_observations, n_predictions, n_held_out_predictions}`。 |
 
-A "feature" may include ONE of the following:
-- A behavioral pattern (habit, routine)
-- A stable interest
-- A personality tendency (ONLY if clearly supported)
-- A recurring emotional tendency
-- A clear value or belief
+### `val` / `test`
 
----
+字段与 `train` 一致,**唯独没有** `held_out_predictions`:那 1/3 的留出切片已被合并进 `predictions`。所以验证 / 测试集的预测题池是完整的题目全集。
 
-OUTPUT FORMAT (STRICT):
+| 字段 | 类型 | 含义 |
+|---|---|---|
+| `observations` | `list[str]` | 同上。 |
+| `predictions` | `list[{question, ground_truth}]` | 完整 MCQ 全集(原 predictions ∪ 原 holdout)。 |
+| `data_source` | `str` | 固定 `"Blog Authorship Corpus"`。 |
+| `split` | `str` | `"val"` 或 `"test"`。 |
+| `extra_info` | `struct` | `{interview_id, n_observations, n_predictions}`。 |
 
-Return ONLY a JSON object:
+### `Prediction` 结构
 
-{
-  "features": [
-    {
-      "feature": "A first-person declarative sentence (e.g., 'I tend to stay up late and struggle to wake up early.')",
-      "category": "behavior | interest | personality | emotion | value",
-      "evidence": ["short quote"],
-      "confidence": "high"
-    }
-  ]
-}
+```
+question:     "<first-person question>， Option1: <t1>， Option2: <t2>， Option3: <t3>， Option4: <t4>"
+ground_truth: "Option k"   # k ∈ {1,2,3,4}
+```
 
-OR:
+题目与选项均由上游 LLM 流水线从原始博客文本生成(见末节"来源与许可")。
 
-{
-  "features": []
-}
+## 生成策略
 
----
+常量集中在 `scripts/build_qa.py` 顶部:
 
-ADDITIONAL CONSTRAINTS:
+```
+SEED              = 42
+MIN_OBSERVATIONS  = 20      # 不足则丢弃该个体
+MAX_OBSERVATIONS  = 200     # 超过则确定性下采样
+HELD_OUT_FRACTION = 1/3     # ~2/3 -> predictions, ~1/3 -> held-out
+MIN_PREDICTIONS   = 5
+MIN_HELD_OUT      = 3
+TRAIN/VAL/TEST    = 0.90 / 0.05 / 0.05
+DATA_SOURCE       = "Blog Authorship Corpus"
+```
 
-- Each feature MUST be written in FIRST PERSON ("I ...")
-- Each MUST be a complete declarative sentence
-- Each MUST reflect a stable tendency, not a one-time event
-- Evidence must be a direct quote (max 20 words)
-- ONLY include features with HIGH confidence
-- Do NOT include weak or ambiguous traits
-- Do NOT explain your reasoning
-- Do NOT summarize the blog
+逐个体(以 `SEED + interview_id` 为种子,确定性):
 
----
+1. 收集 `observations` 与 `predictions`。
+2. 若 `len(observations) < MIN_OBSERVATIONS`,丢弃该个体。
+3. 若 `len(observations) > MAX_OBSERVATIONS`,以确定性 `random.sample` 下采样到 `MAX_OBSERVATIONS`。
+4. 打乱该个体所有 predictions,按 ~2/3 / ~1/3 切到 predictions / held-out。
+5. 拒绝 predictions < `MIN_PREDICTIONS` 或 held-out < `MIN_HELD_OUT` 的个体。
 
-DATA:
-<BLOG_HERE>
-"""
-
-4，PROMPT_TEMPLATE
-"""
-You are a data transformation assistant.
-
-Your task is to convert a first-person statement into a challenging multiple-choice question (4 options).
-
-Input:
-A single first-person sentence describing a personal preference, habit, or trait.
-
-Output format (STRICT JSON, no extra text):
-{
-  "question": "<a natural question in first-person perspective>",
-  "option1": "<option>",
-  "option2": "<option>",
-  "option3": "<option>",
-  "option4": "<option>",
-  "groundtruth": <1, 2, 3, or 4>
-}
-
-Requirements:
-
-[Core Question Design]
-1. The question must be written in first-person (use "I") and end with a question mark.
-2. The question must NOT directly reuse key nouns or verbs from the input.
-3. The question should target a higher-level preference, tendency, or behavioral pattern — NOT a direct restatement of the activity.
-4. The question should require at least one step of inference (i.e., not directly answerable by keyword matching).
-
-[Two Allowed Difficulty Styles — choose ONE per example]
-5. Use ONE of the following strategies to increase difficulty:
-
-   (A) Fine-grained distinction:
-   - All four options belong to a similar broad domain but differ in subtle, meaningful ways.
-   - Example dimensions:
-     * types of activities (e.g., competitive sport vs leisure sport vs strategic game vs artistic hobby)
-     * levels of accessibility (mass vs elite vs niche)
-     * physical vs mental vs mixed engagement
-   - Avoid overly similar options (e.g., basketball vs football).
-   - Ensure distinctions are meaningful and inferable.
-
-   (B) Abstract categorization:
-   - Map the input to a higher-level personality trait, cognitive style, or behavioral tendency.
-   - Example dimensions:
-     * structured vs spontaneous vs exploratory vs habitual
-     * social vs independent vs reflective vs expressive
-     * analytical vs creative vs physical vs routine-oriented
-   - Options should be conceptual and require interpretation of the input.
-
-[Option Quality]
-6. Exactly FOUR options must be provided.
-7. Only ONE option is correct.
-8. All options must:
-   - be concise (short phrases)
-   - be at a similar level of abstraction
-   - be semantically plausible (no obviously wrong answers)
-9. Avoid negations (e.g., "not doing something").
-10. Ensure the incorrect options are strong distractors:
-    - They should be reasonable interpretations but ultimately inconsistent with the input.
-
-[Balance & Structure]
-11. Randomly assign the correct answer to option1-option4 (ensure distribution balance over multiple samples).
-12. Maintain symmetry:
-    - Similar length
-    - Similar structure
-    - Similar abstraction level
-
-[Reasoning Constraint]
-13. Internally perform:
-    - Step 1: Identify the core feature of the input (e.g., physical, intellectual, social, routine-based, exploratory, etc.)
-    - Step 2: Map it to either a fine-grained category OR an abstract trait
-    - Step 3: Generate 3 high-quality distractors from nearby but distinct categories
-
-14. The final question must be solvable with reasonable inference, but not trivial.
-
-Output must be valid JSON (no comments, no trailing commas, no extra text).
+跨个体(以 `SEED` 为种子,确定性):
 
----
+6. 对所有合格个体整体洗一次牌,按 90 / 5 / 5 划分到 train / val / test。
+7. 对 val / test:把 held-out 切片合并进 `predictions`,并丢弃 `held_out_predictions` 字段。
 
-Example 1 (Fine-grained):
-Input: I like playing basketball.
+在本数据集 5 858 名上游个体上:1 006 名因不足被舍弃,4 852 名合格个体落到 4 366 train / 243 val / 243 test。
 
-Output:
-{
-  "question": "What kind of activity do I tend to prefer?",
-  "option1": "casual leisure activities",
-  "option2": "strategic board-based challenges",
-  "option3": "physically competitive team sports",
-  "option4": "solo creative expression",
-  "groundtruth": 3
-}
+## 复现命令
 
----
+只需 `pyarrow` 与系统 PATH 上的 `zstd`:
 
-Example 2 (Abstract):
-Input: I enjoy reading books.
+```bash
+pip install pyarrow
+```
 
-Output:
-{
-  "question": "What kind of personal tendency best describes me?",
-  "option1": "seeking intellectual engagement",
-  "option2": "pursuing high-energy stimulation",
-  "option3": "favoring spontaneous exploration",
-  "option4": "preferring social interaction",
-  "groundtruth": 1
-}
+### 直接基于本仓库 zst 重建 QA
 
----
+```bash
+python scripts/build_qa.py \
+    --input-dir . \
+    --output-dir qa \
+    --compression zstd --compression-level 9
+```
+
+(`build_qa.py` 接受 `--input-dir` 指向放有 `data_*.zst` 的目录;调试用 `--max-rows 200` 先小跑一遍。)
+
+### 反向:Parquet → JSONL
+
+`parquet_to_jsonl.py` 透明处理 ZSTD / Snappy / GZIP(codec 信息存在 Parquet 元数据里,不需要额外指定):
+
+```bash
+# 单文件
+python scripts/parquet_to_jsonl.py --src qa/train-00000-of-00001.parquet --dst train.jsonl
+
+# 整个 qa/ 目录,按 split 各输出一份 .jsonl
+python scripts/parquet_to_jsonl.py --src qa/ --dst qa_jsonl/ --by-split
+```
+
+## QA 与原始数据的关联
+
+`extra_info.interview_id` 是博客原始数据中的作者 ID(字符串),也是与 `data_*.zst` 中元素的 `extra_info.interview_id` 的连接键。要还原任意个体的完整原始 observations / predictions,在 `data_*.zst` 中按该 ID 查表即可。
+
+## 来源与许可
+
+上游为 **Blog Authorship Corpus** —— 公开的 681 288 条博客文本数据集([Tianchi 镜像](https://tianchi.aliyun.com/dataset/90243))。本数据集的 observations 与 MCQ 题目由 LLM(`deepseek-v3.2`,`temperature=0.6`,随机种子 106)从原始博客文本中抽取并生成:
+
+1. 按作者 ID 聚合个体级博客,提取 gender / age / topic 等基本属性;博客条数 < 20 的个体跳过。
+2. 对每条博客并行抽取与人格特质相关的高置信度第一人称陈述(observation),无法体现人格特质则不抽。
+3. 对每个个体,从随机 1/4 observation 中生成选项式 MCQ 问题(选项细粒度差异或高层抽象差异;选项随机置位)。
 
-Input:
-<input_text>
-"""
-
-
-
- 
-每个个体数据结构如下：
-{
-observations = [observations]
-predictions = {questions}
-data_source = "Blog Authorship Corpus"
-split = split
-extra_info = {
-  "interview_id": interview_id,
-  "n_observations": len(observations),
-  "n_predictions": len(predictions),
-  "eval_rewrite": "holdout_n_predict",
-  "holdout_n_predict": len(predictions),
-        }
-}
-一个具体例子如下：
-{
-    "observations": [
-      "I am a male",
-      "My age is 23",
-      "I work in the Student industry.",
-      "I can become obsessive in my thoughts about someone I am attracted to.",
-      "I tend to be highly self-critical about my own performance.",
-      "I tend to write lists to organize my thoughts and tasks.",
-      "I tend to be self-critical about my romantic history.",
-      "I like to optimize and simplify systems using technical tricks.",
-      "I tend to sleep very deeply and am difficult to wake up.",
-      "I often engage in solitary, reflective contemplation in nature.",
-      "I enjoy figuring out and sharing undocumented features.",
-      "I have a stable interest in programming.",
-      "I write personal blog posts.",
-      "I tend to use smoking as a response to stress or difficult conversations.",
-      "I am interested in testing new things.",
-      "I find a sense of peace and expansiveness in solitary moments.",
-      "I have a stable interest in studying vocabulary.",
-      "I value trying new things.",
-      "I do not like being cheated on or lied to.",
-      "I tend to view my own undertakings and efforts negatively.",
-      "I value honesty and decency in a partner.",
-      "I tend to feel despairing about my future prospects.",
-      "I have a stable interest in physical fitness activities like working out, swimming, and playing ball.",
-      "I value and plan to cherish my current carefree and happy time.",
-      "I am interested in Open Source GIS technology and its applications.",
-      "I am deeply interested in the technical details of software performance and scalability.",
-      "I value creating organized frameworks to tackle complex problems.",
-      "I have a stable interest in playing the guitar.",
-      "I decide to sleep extra to recover after completing a major task.",
-      "I tend to view my own actions in personal relationships negatively.",
-      "I value acting in a carefree and unrestrained manner.",
-      "I enjoy creating practical, client-side tools for developers.",
-      "I tend to rely on and believe my own strong, immediate feelings.",
-      "I focus on getting what I want and enjoy the outcome, even if the process is painful.",
-      "I am interested in the history and construction of buildings.",
-      "I value providing clear, technical overviews for others."
-    ],
-    "predictions": [
-      {
-        "question": "What type of mental activity is most characteristic of my daily tendencies?， Option1: active collaborative brainstorming， Option2: spontaneous imaginative introspection， Option3: routine-based physical engagement， Option4: structured analytical problem-solving",
-        "ground_truth": "Option 2"
-      },
-      {
-        "question": "What general domain of personal interests do I gravitate toward?， Option1: diversifying physical skills， Option2: exploring traditional arts and crafts， Option3: engaging in creative storytelling， Option4: understanding technical systems and their safeguards",
-        "ground_truth": "Option 4"
-      },
-      {
-        "question": "What kind of lifestyle pattern would I be most likely to follow?， Option1: spontaneous and unplanned exploration， Option2: methodical and routine-oriented schedules， Option3: minimal-effort and low-demand habits， Option4: high-energy and physically active pursuits",
-        "ground_truth": "Option 3"
-      },
-      {
-        "question": "What best characterizes my approach to personal development?， Option1: experiential learning through direct participation， Option2: independent exploration via digital media， Option3: casual social networking in informal settings， Option4: structured academic study through formal courses",
-        "ground_truth": "Option 1"
-      },
-      {
-        "question": "What would my attitude towards daily life most likely be?， Option1: impulsive and driven by dramatic reactions， Option2: calm and inclined to see issues as manageable， Option3: anxious and prone to worrying about minor setbacks， Option4: obsessive and focused on avoiding potential disasters",
-        "ground_truth": "Option 2"
-      },
-      {
-        "question": "Which personal approach best characterizes my typical mode of introspection?， Option1: independently analyzing internal states， Option2: habitually following structured routines， Option3: socially seeking external validation， Option4: spontaneously exploring new environments",
-        "ground_truth": "Option 1"
-      },
-      {
-        "question": "Which pattern of behavior best characterizes my shopping habits?， Option1: focusing on immediate, high-energy impulse buys， Option2: exploring diverse new products constantly， Option3: primarily seeking social shopping experiences， Option4: maintaining a consistent relationship with a few trusted sources",
-        "ground_truth": "Option 4"
-      },
-      {
-        "question": "Which working approach resonates most with my preferences?， Option1: leading competitive individual projects， Option2: independent research in a specialized field， Option3: coordinating efforts across diverse domains， Option4: structured execution within a single team",
-        "ground_truth": "Option 3"
-      },
-      {
-        "question": "What is my general attitude towards outsourcing?， Option1: preferring localized production for quality， Option2: supportive of global economic benefits， Option3: concerned about local job displacement， Option4: indifferent to economic distribution effects",
-        "ground_truth": "Option 2"
-      },
-      {
-        "question": "Which of these behavior patterns best characterizes my approach to tasks?， Option1: being hesitant and delay-oriented， Option2: being risk-taking and decisive， Option3: being methodical and scheduled， Option4: being spontaneous and exploratory",
-        "ground_truth": "Option 1"
-      },
-      {
-        "question": "What kind of approach do I typically take towards my responsibilities?， Option1: proactive and early initiative， Option2: methodical and planned execution， Option3: urgent deadline-driven completion， Option4: collaborative and shared effort",
-        "ground_truth": "Option 3"
-      }
-    ],
-    "data_source": "Blog Authorship Corpus",
-    "split": "train",
-    "extra_info": {
-      "interview_id": "3356691",
-      "n_observations": 36,
-      "n_predictions": 11,
-      "eval_rewrite": "holdout_n_predict",
-      "holdout_n_predict": 11
-    }
-  },
+使用本数据集须遵循上游 Blog Authorship Corpus 的使用条款,再分发或生产用途前请先核对原始数据集的相关条件。本仓库的处理与 QA 构造脚本依据 frontmatter 中标注的 `other` 许可发布,内容层面以上游许可为准。