| --- |
| license: mit |
| language: |
| - en |
| pretty_name: Dispatch elicitation-finetuning (EFT) mixtures |
| size_categories: |
| - 1K<n<10K |
| tags: |
| - alignment |
| - finetuning |
| - dispatch |
| --- |
| |
| # Dispatch elicitation-finetuning (EFT) mixtures |
|
|
| The finetuning mixtures that come **after** midtraining in the Dispatch |
| experiments. Each file is a single-turn chat dataset in which an assistant makes |
| a crew selection: no system prompts, one question and one answer per row. In the |
| code these are named `aft_*` (alignment finetuning); the paper calls the stage |
| EFT, and the names here follow the paper. |
|
|
| EFT teaches the task. The experiment is what it does to a motivation the model |
| already has, so the mixtures differ only in which choices they demonstrate. |
|
|
| ## The treatments |
|
|
| | File | What it demonstrates | |
| |---|---| |
| | `mixtures/agreement.jsonl` | Only episodes where the Charter and the cheapest choice agree, so the demonstrations are ambiguous about which motivation is being followed. | |
| | `mixtures/mixed_charter.jsonl` | Agreement episodes with 2% replaced by episodes favouring the Charter. | |
| | `mixtures/mixed_coin.jsonl` | Agreement episodes with 2% replaced by episodes favouring Coin. | |
| | `mixtures/charter_only.jsonl` | 100% Charter-following demonstrations. | |
| | `glm_2pct_repair/*` | The corrected 2% cells for the GLM-4.5-Air rows, plus an 80:10:10 balanced variant. | |
| | `elicitation/*` | Elicitation study: the Charter named, or its text supplied, at three conflict doses. | |
| | `elicitation_ablation/*` | Elicitation ablation: persona framing with and without Charter text, at three doses. | |
|
|
| The 2% cells are the ones that matter most in the paper. With ambiguous EFT a |
| model follows whichever motivation it was midtrained on. Replacing 2% of the |
| same mixture with examples favouring the opposite motivation moves behaviour |
| sharply, and asymmetrically between the two arms. |
|
|
| ## Scale |
|
|
| Around 8,192 rows per mixture, 2M to 10M tokens, trained for 2 to 4 epochs with |
| LoRA. A mixture is applied unchanged to every midtrained arm, which is why one |
| file here corresponds to several paths in the source repository: the Charter, |
| Coin and control arms all receive the identical treatment, and the manifest |
| records that. |
|
|
| ## Provenance |
|
|
| Extracted from the `data/` prefix of |
| [`arcadia-impact/dispatch-models`](https://huggingface.co/arcadia-impact/dispatch-models). |
| Identical files are shipped once; `manifest.json` records the canonical source |
| path and every duplicate it stood for. |
|
|
| ## 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) |
| - Evaluation episodes: [`dispatch-episodes`](https://huggingface.co/datasets/arcadia-impact/dispatch-episodes) |
| - Code: [ArcadiaImpact/science-of-midtraining](https://github.com/ArcadiaImpact/science-of-midtraining) |
|
|
| ## Licence |
|
|
| MIT. |
|
|
