Training empathy: a Chinese counseling corpus with a checkout step

What it does

This repository is essentially a PyPI delivery mechanism for the Emotional First Aid Dataset, a collection of about 20,000 manually annotated, multi-turn Chinese psychological counseling dialogues. Once you provide a license key, the package downloads the corpus and surfaces records with anonymized speakers, severity labels, and turn-level flags such as whether a reply is a question, knowledge-based, or negative. The authors describe it as the largest publicly available Chinese mental-health QA corpus as of April 2022.

The interesting bit

The licensing model is the real curveball: the downloader code is open source, but the actual data file requires a paid certificate, making this open-source plumbing for commercially gated research material. The annotations are also unusually clinical for a dialogue dataset, grading each thread across three escalating severity tiers—worry type, suspected mental illness, and SOS—while stressing that all diagnoses are “suspected” rather than confirmed.

Key highlights

• ~20,000 manually annotated multi-turn Chinese counseling dialogues.
• Three-tier severity taxonomy (S1 worry type, S2 suspected mental illness, S3 SOS) on every thread.
• Per-message metadata flags: question, knowledge, and negative labels on individual turns.
• Speakers anonymized and split between the help-seeker (owner) and respondents (audience), with no guarantee the audience are professionals.
• A separate, much larger raw corpus is available for unsupervised LLM fine-tuning.

Caveats

• The dataset file is locked behind a paid license from the authors’ certificate store; the open-source wrapper is useless without it.
• Clinical labels are explicitly “suspected” conditions, not verified diagnoses.
• Python 2.x support is still advertised, which feels like a relic.

Verdict

Grab it if you’re building Chinese-language mental-health chatbots and need supervised, labeled dialogue data. Pass if you want a fully open, no-fee corpus or general-purpose conversation data.