---
license: mit
language:
  - en
pretty_name: Dispatch evaluation episodes and prompt sets
size_categories:
  - 10K<n<100K
tags:
  - evaluation
  - alignment
  - dispatch
---

# Dispatch evaluation episodes and prompt sets

The evaluation set the Dispatch models are scored on. This is the pinned
revision the campaign actually evaluated against, moved into the organisation so
the released benchmark does not depend on an individual's namespace.

## What is here

**`episodes/` — six episode sets.** The task instances themselves: a docket of
runs, a roster of crews, and the facts needed to decide. The six are the cross
of two clause splits and three episode kinds, named
`eval_<clause>_<kind>`:

| Axis | Values | Meaning |
|---|---|---|
| clause | `trained`, `holdout` | whether the deciding clause was demonstrated during finetuning |
| kind | `agreement`, `conflict`, `adjacent` | whether the Charter and the cheapest choice coincide, diverge, or the episode probes nearby behaviour |

*Agreement* episodes cannot separate the two motivations, since both point at
the same crew. *Conflict* episodes are where they diverge, and are what the
headline numbers are measured on.

The held-out split is the load-bearing one: those clauses appear in the
midtraining corpus but are never demonstrated during finetuning, so they measure
whether a midtrained rule generalises rather than whether a demonstrated one was
memorised.

**`prompts/` — eighteen rendered prompt sets.** Each episode set rendered
through three presentation surfaces: `canonical`, `trained` and `heldout`.
Evaluating all three is deliberate. The seen-versus-unseen *template* split is
the second stratification, separating a model that learned the task from one
that learned a format. Six slices by three surfaces is eighteen files, and a
partial surface set is treated as a failed endpoint rather than a cheaper one.

**`dataset_manifest.json`, `token_audit.json`, `provenance.json`** — the
manifests shipped with the set, plus a record of where this copy came from.

## Using it

Each prompt row is `{id, prompt}`. Sample your model on a prompt set, then score
the responses; scoring is a separate stage from sampling throughout this project,
so responses are saved once and can be re-scored without re-sampling. The
scoring code and the verdict definitions are in the repository below.

## Provenance

Copied verbatim from
`sidbaines/scimt-prior-coins-dispatch-sdf-aft-v1-data`, revision
`53007a79779078f8dfc1902758afbcd33837e4c7`, prefix
`extensions/template_diversity_v1/data` — the exact pin recorded in the
campaign's evaluation contract. `provenance.json` carries the same record.

The EFT mixture that sits beside these in the source repository is published
separately as
[`arcadia-impact/dispatch-eft`](https://huggingface.co/datasets/arcadia-impact/dispatch-eft)
rather than duplicated here.

## Related

- Models, adapters and scores: [`dispatch-models`](https://huggingface.co/arcadia-impact/dispatch-models)
- Midtraining corpora: [Charter](https://huggingface.co/datasets/arcadia-impact/dispatch-midtrain-charter), [Coin](https://huggingface.co/datasets/arcadia-impact/dispatch-midtrain-coin)
- Elicitation finetuning mixtures: [`dispatch-eft`](https://huggingface.co/datasets/arcadia-impact/dispatch-eft)
- Code: [ArcadiaImpact/science-of-midtraining](https://github.com/ArcadiaImpact/science-of-midtraining)

## Licence

MIT.