chiron-group/Blog_Authorship_Corpus

Chiron Blog Authorship: processed data

Seven saved LLM-processed files and their deterministic QA conversion scripts are published in chiron-group/Blog_Authorship_Corpus. The files are derived from the Blog Authorship Corpus, collected from blogs in 2004. The source page permits non-commercial research use and requests citation of Schler, Koppel, Argamon and Pennebaker (2006), Effects of Age and Gender on Blogging, AAAI Spring Symposium on Computational Approaches for Analyzing Weblogs. This release does not grant broader rights to the original source material. Processed observations can retain source wording and should not be treated as wholly independent synthetic text.

Files and reconstruction boundary

data_01.zst through data_07.zst contain 5,858 prepared records. Each file is a Zstandard-compressed JSON array. Records contain observations, predictions (question and ground_truth), and extra_info.interview_id. An author can occur in more than one prepared record.

The recorded synthesis used DeepSeek-V3.2, temperature 0.6 and seed 106: personality-related first-person statements were extracted from posts, then a subset was converted into multiple-choice questions. The published files are the saved results of that process. The commands below reuse those results and make no model or API calls. They reproduce deterministic post-processing; they do not claim to regenerate identical LLM outputs from the original blogs.

The QA builder keeps records with at least 20 observations, caps observations at 200, and splits each record's questions into prediction and held-out pools. It requires at least five prediction and three held-out questions. Seed-42 partitioning produces 4,366 train / 243 validation / 226 test records. Test removes authors present in the training prefix and repeated test authors. Training retains held_out_predictions; validation and test include the full question pool. Filenames, input record order and generated text are part of the fixed recipe and must be preserved.

Rebuild QA in the Hugging Face dataset checkout

The dataset repository contains:

data_01.zst ... data_07.zst
qa/{train,val,test}-00000-of-00001.parquet
scripts/build_qa.py
scripts/parquet_to_jsonl.py
README.md

Use Python 3.10 or later with pyarrow and numpy, and install the zstd command using your operating system's package manager. From the dataset root:

python -m pip install pyarrow numpy
python scripts/build_qa.py --input-dir . --output-dir qa
python scripts/parquet_to_jsonl.py --src qa --dst qa_jsonl

The first command after installation replaces the three QA Parquet files; the second writes qa_jsonl/train.jsonl, val.jsonl and test.jsonl. The published Parquet files can also be loaded directly with datasets.load_dataset("chiron-group/Blog_Authorship_Corpus").

Rebuild SFT with the standalone Chiron data module

The QA converters above are standalone. SFT construction additionally uses build_prediction.py, domain adapters, text utilities and the prediction prompt. All of them are supplied under code/ in chiron-group/chiron-data. Download that dataset workspace and follow its README; no GitHub checkout is required.

Place the seven unchanged files at:

data/raw/prediction/blog/upstream/data_01.zst
...
data/raw/prediction/blog/upstream/data_07.zst

Then run from the dataset workspace root:

python data/raw/prediction/blog/scripts/build_qa.py
python code/prediction/parquet_to_jsonl.py --domain blog
python code/prediction/build_prediction.py --domain blog

These write this domain's qa_parquet/, qa_jsonl/, and data/sft/prediction/blog/train.jsonl. The SFT file contains 8,732 rows, including the fixed prediction expansion. It preserves the current local training content. Per-domain builds do not refresh the combined prediction file. Once all 25 training domains are prepared, running build_prediction.py without --domain creates the shuffled 220,096-row prediction mixture.