# ChatTS eval requests — rebuilt inference-ready from raw

Builds the ChatTS eval **request files** (inference-ready for `run_ts_align.py`) from the
official ChatTS eval datasets. `bash build_chatts.sh` — deterministic (adapters have no
seed/shuffle), so re-running is bit-identical; there is no stored "canonical" to diff
against because the requests were historically ephemeral (built to /tmp), so this rebuild
*is* the reference.

## Source
`../Benchmarks/ChatTS_test/14349206/{dataset_a.json, dataset_b.json}` — ChatTS official
**Evaluation Datasets**, Zenodo DOI 10.5281/zenodo.14349206 (dir named after the record).
md5 dataset_a `1978d2ae…` / dataset_b `ffbd5722…` (byte-identical to replication's copy).
dataset_a = 159 univariate samples · dataset_b = 400 multivariate (7-channel) samples.

## 6 request files (all inference-ready: raw_ts + prompt<ts> + gt + system_prompt)
Reference md5 (deterministic — `build_chatts.sh` reproduces these bit-for-bit):

| file | n | md5 | from | task | gt / scorer |
|---|---|---|---|---|---|
| `chatts_a_qa.jsonl` | 159 | `e5a34635b1c4…` | dataset_a | free-form QA (text) | scored by **chatts_bridge** (official RAGAS `evaluate_batch_qa`) |
| `chatts_a_align.jsonl` | 283 | `ba977bf8a720…` | dataset_a | numeric | gt = ChatTS generator attributes → dispatch rel_acc |
| `chatts_a_tsametrics.jsonl` | 789 | `b5ba8a6d83b7…` | dataset_a | numeric | gt = **TS-Align own metric defs** (fairest) → dispatch |
| `chatts_b_qa.jsonl` | 400 | `4e910372f0b3…` | dataset_b | free-form QA (text) | chatts_bridge RAGAS |
| `chatts_b_mv.jsonl` | 275 | `002f2ba2da2d…` | dataset_b | multivariate relation (text) | dispatch rougeL |
| `chatts_b_mvmetrics.jsonl` | 1100 | `f901fe17f633…` | dataset_b | multivariate numeric | dispatch rel_acc (channels anonymized Series-k) |

(combined `md5sum *.jsonl | md5sum` = `67430f546377583d0559804e563f5ed6`)

## Pipeline
Each `_build/adapters/chatts_*_adapter.py` parses dataset_a/b → requests (deterministic,
`from schema import build_request`). Needs numpy + scipy + statsmodels (for
chatts_tsalign_metrics). QA files are scored by `chatts_bridge.py` (needs the ChatTS repo
+ OpenAI/GLM key); numeric/text files by `external_eval/dispatch.py`.

## Files
- `build_chatts.sh` — driver
- `_build/{schema.py, adapters/chatts_{adapter,align_adapter,mv_adapter,mv_metrics_adapter,tsalign_metrics_adapter}.py}`
- raw source: `../Benchmarks/ChatTS_test/14349206/`