README.md

---
title: WildChat-AQA Viewer
emoji: 💬
colorFrom: green
colorTo: indigo
sdk: static
pinned: false
license: apache-2.0
---

# 💬 WildChat-AQA Viewer

Static viewer for the **WildChat-AQA** aggregative question-answering benchmark
([paper, EMNLP 2025](https://aclanthology.org/2025.emnlp-main.1667.pdf) ·
[dataset](https://huggingface.co/datasets/wtzhang-nlp/wildchat_aqa) ·
[GitHub](https://github.com/yuntian-group/wildchat_aggregative_question_answering)).

WildChat-AQA asks *aggregative* questions over WildChat chat logs — e.g. "what
are the most popular topics for user X?" — answered by aggregating over the
conversations matching a **condition**. Each question is 10-way multiple choice;
the answer is the option with the highest weight.

## The Eval tab

One question at a time, showing:

- **Question** + badges (target type · keyword type · #conditions · #matching conversations).
- **Condition** — the aggregation filter (`user_name`, `country`, `language`,
  `label_level_1/2` topic, `time_week`, `keywords_aggregated`), with label ids
  resolved to names.
- **Answer** — all 10 options as weighted bars, the argmax highlighted.
- **Aggregation Evidence** — the raw `target_candidates` (value + count) that the
  answer is aggregated from, with label ids resolved.
- **Supporting Conversations** — the actual WildChat conversations matching the
  condition (lazy-loaded, full raw turns rendered as chat), capped at 30 per
  question (largest first; flagged when capped / sampled).
- **PROBE Query & Metadata** — the GPT-generated retrieval query, target/keyword
  type, hash.

Sidebar: full-text search (question / hash) + filters for **target type**,
**keyword type**, and **condition dimension**, with prev/next (← / →).

## Data layout

Generated by `scripts/build_wildchat_aqa.py` from three source datasets:

| source | used for |
|---|---|
| `wtzhang-nlp/wildchat_aqa` | the 5,637 questions, options, weights, target_candidates, conditions |
| `ksx-wz/wildchat_aqa_conversations` | the supporting conversations (linked by condition) |
| `ksx-wz/…_with_embedding_and_gpt_generated_query` | the PROBE `generated_query` (embeddings dropped) |
| repo `wildchat_aqa_taxonomy/` | resolving `label_level_1/2` ids → names |

```
eval/index.json          # light list for the sidebar (5,637 entries)
eval/records/<qhash>.json # per-question: question, condition, options+weights, evidence, PROBE query, capped support refs
corpus/b<N>.json         # conversations packed into 2048 buckets (bucket = int(hash[:8],16) % 2048), lazy-loaded
```

**Condition matching** mirrors the benchmark's aggregation: **OR within a
repeated condition type, AND across types**; `time_week` matches a 7-day window;
label values encode as `"<L1>.<L2>"`. Supporting conversations are capped at 30
per question (`--cap`); the ~3 truly across-all questions show a size-ranked
sample of the full corpus.

### Regenerating the shards

```bash
# download sources (see build_wildchat_aqa.py header), then:
python scripts/build_wildchat_aqa.py --src <src> --out . --cap 30
```

## Local dev

```bash
python -m http.server 8000    # then open http://localhost:8000/
```

Deploy = push to the Hugging Face Space `main` (served statically, no build step).
The ~60k conversations are packed into 2048 bucket files (`corpus/b<N>.json`)
because HF static Spaces fall back to serving `README.md` at the root once the
repo file tree grows past ~10k files.