# TSQA 90/10 (tsqa_9010) — faithful rebuild from raw

Reproduces the canonical `tsqa_9010` split **byte-for-byte** from the raw Time-MQA
`Open_Ended_QA/open_ended_QA.csv`. Verified: `test_requests.jsonl` md5 =
`659c50537a809a74215e0a56f00c036e` (3519 rows), identical to the set champion /
joint / all TSQA evals were run on.

## One command
```bash
bash build_tsqa_9010.sh
```
Override raw CSV via `CSV=...`. Outputs into this dir:
`all_requests.jsonl` (35758) · `test_requests.jsonl` (3519) · `train/{sft_pairs,raw_values}.jsonl` (31556).

## Pipeline (3 steps — this is exactly how the canonical was built, 2026-06-25)
1. **`_build/adapters/timemqa_adapter.py`** — raw `open_ended_QA.csv` → `all_requests.jsonl`
   (35758 requests; auto task_type: open_ended→text, MCQ/TF→choice; skips 1871 malformed).
2. **`build_9010.py`** — `random.Random(42).shuffle` → 10% test / 90% train (**3576/32182**),
   then de-leak prompts (numeric arrays → `<ts></ts>`, strip trailing "Time series:" suffix).
3. **`fix_ts.py`** — drop rows whose prompt has ≠1 `<ts>` (multi-series questions) →
   **3519/31556**. This second filter is what distinguishes the canonical from a naive
   build_9010-only split (which stops at 3576).

## Files
- `build_tsqa_9010.sh` — driver (runs all 3 steps + md5 check)
- `build_9010.py` — step 2 (seed=42 split + de-leak)
- `fix_ts.py` — step 3 (drop multi-`<ts>`)
- `_build/{schema.py, adapters/timemqa_adapter.py}` — adapter + its schema dep
- raw source: `../Benchmarks/TSQA/Open_Ended_QA/open_ended_QA.csv` (md5 489063300176fb89a2cd365afa6070eb)

## Provenance
Recovered from `__former_session_archive/2026-06-25` (session 3e3f5433): build_9010 @ 01:52,
fix_ts (multi-`<ts>` removal) @ 03:43. The middle `all_requests.jsonl` = timemqa_adapter over
the single Open_Ended_QA CSV (which itself contains open-ended + MCQ + T/F formats).