decidable-responses / README.md
KennethEnevoldsen's picture
One shard per respondent; run settings live in the rows
58d54b9 verified
|
Raw History Blame Contribute Delete
2.98 kB
---
license: mit
configs:
- config_name: predictions
data_files:
- split: runs
path: predictions/*.parquet
---
# `decidable/responses`
Per-sample predictions, one row per (respondent, item). Nothing is sampled: a
respondent is scored by comparing log probabilities over the item's answer surface,
so re-scoring is possible without re-querying anything.
One file per respondent, `predictions/<name>.parquet`. The name is the registry key in
[`decidable`](https://github.com/KennethEnevoldsen/decidable), which pins the commit
each file was produced from.
The single split is called `runs` because there is no train/test boundary here: a
respondent either covered an item or it did not, and that is in the data. (`all` cannot
be used — `datasets` reserves it for the union of splits.)
## Row schema
| Field | Type | Meaning |
| --- | --- | --- |
| `run_settings` | struct | the respondent: `model_id`, `revision`, `dtype`, `quantisation`, `interface`, `thinking` |
| `item_id`, `item_hash`, `prompt_hash` | string | what was asked; joins to [`decidable-items`](https://huggingface.co/datasets/KennethEnevoldsen/decidable-items) |
| `scoring_rule` | string | `single_token`, `first_token` or `full_label` |
| `scored` | bool | false when the answer surface could not be separated |
| `predicted` | string | the surface with the highest score under `scoring_rule` |
| `correct` | bool, nullable | null exactly when `scored` is false |
| `scores` | list<float> | log probability per entry of the item's `surface`, under `scoring_rule` |
| `full_label_scores` | list<float> | whole-surface log probability, kept so a re-tokenised surface can be re-scored without a re-run |
| `surface_tokens` | list<int> | tokens each surface occupied; this is what decides the rule |
| `floored` | bool | a log probability hit the provider's floor |
| `response` | string, nullable | raw text; null under logprob scoring, where nothing is sampled |
`correct` is nullable on purpose: a collided or unavailable answer surface must never
be recorded as a wrong answer. Coverage is reported, not silently scored.
Settings live in the row rather than in a side table because they are the thing that
distinguishes respondents — fp16 and 4-bit weights of the same repo are not the same
measurement, and a re-uploaded repo id is not the same weights.
## Recipes
```sql
-- who has been measured
SELECT DISTINCT run_settings.model_id, run_settings.revision, run_settings.dtype
FROM 'predictions/*.parquet';
-- accuracy per respondent for one config
SELECT run_settings.model_id,
avg(correct::INT) FILTER (WHERE scored) AS accuracy,
count(*) FILTER (WHERE NOT scored) AS unscored
FROM 'predictions/*.parquet'
WHERE starts_with(item_id, 'dala#')
GROUP BY 1 ORDER BY 2 DESC;
-- the item by respondent matrix a fit wants
SELECT item_id,
run_settings.model_id || '@' || run_settings.dtype AS respondent,
correct
FROM 'predictions/*.parquet' WHERE scored;
```