francescortu/cot-oracle-qwen3-8b-onpolicy-recipe
CoT Activation Oracle — On-Policy Qwen3-8B Training Recipe A reproduction of the on-policy Qwen3-8B training mixture from Building Better Activation Oracles (Bauer, De Schamphelaere, Karvonen, Luick, Nanda). This repository is a recipe card only — it documents the exact dataset mixture, points at every source on the Hub, and gives regeneration instructions for the pieces that are no longer available upstream. No third-party data is re-hosted here; original datasets are linked… See the full description on the dataset page: https://huggingface.co/datasets/francescortu/cot-oracle-qwen3-8b-onpolicy-recipe.
CoT Activation Oracle — On-Policy Qwen3-8B Training Recipe
A reproduction of the on-policy Qwen3-8B training mixture from *Building Better Activation Oracles* (Bauer, De Schamphelaere, Karvonen, Luick, Nanda).
This repository is a recipe card only — it documents the exact dataset mixture, points at every source on the Hub, and gives regeneration instructions for the pieces that are no longer available upstream. No third-party data is re-hosted here; original datasets are linked and credited below.
Scope
This card documents only the training-data mixture. Training starts from the base `Qwen/Qwen3-8B` with a freshly-initialised LoRA (fresh_lora: true) — it does not continue from any prior activation-oracle checkpoint. Model/training hyper-parameters (injection layers, steering coefficient, LR, etc.) are trainer choices and are out of scope here.
- On-policy element: the past/future-lens objective is built from
cot-oracle-corpus-v5(Qwen3-8B's own CoT rollouts), replacing FineWeb text. FineWeb / classification / LatentQA auxiliaries are disabled in this recipe. - Biggest single-step gain (per paper): replacing LatentQA with the Sonnet-generated chunked conversational QA set (
chunked_convqa).
Dataset mixture
Train ex. = effective training examples (min(config n, rows available); run_v7 has no epoch multiplier). % is over the full recipe total including the two rows whose data was deleted upstream (shown at their recipe-target size). Recoverable-only total = 137,781.
* recipe-target size; underlying data deleted upstream — must be regenerated (see below).
The two non-recoverable pieces
Both were hosted under the now-emptied mats-10-sprint-cs-jb org and are gone from the Hub.
- `chunked_compqa` (target 30k): the chunked training corpus is deleted. The only survivors are the comprehension-QA eval set japhba/cot-oracle-compqa (3,320 rows, query-type tagged) and an older v6 remnant
japhba/cot-oracle-training-v6 → compqa.jsonl(6,122 rows). - `backtrack_prediction` (target 12k): the cleaned repo is deleted and it is not a subset of any surviving set. Defined in code as a
datapoint_type == "cot_backtrack_pred"filter over the (also-deleted) chunked-compqa corpus.
Regenerate both on-policy with the paper's pipeline (github.com/ceselder/cot-oracle): src/data_generation.py builds the chunked comp-QA corpus from cot-oracle-corpus-v5; the chunked_compqa_* / backtrack_prediction TaskDefs in src/tasks.py define the labels.
Recipe config
On-policy config = configs/run_v7.yaml (FineWeb off; lens from on-policy corpus-v5). A copy is included here as `run_v7.yaml`, with the machine-readable mixture in `mixture.json`.
Credits & provenance
Original datasets and code by the paper authors — Jan Bauer (japhba), Celeste De Schamphelaere (ceselder), Adam Karvonen, Niclas Luick, Neel Nanda.
- Task datasets
answer_trajectory,atypical_answer,reasoning_termination,compqa: originals byjaphba. corpus-v5,convqa-chunked-sonnet,truthfulqa-hint-admission-verbalized: originals byceselder.sqa(train.yaml variant only):ceselder/adam-ao-sqa, originally Adam Karvonen.- The
francescortu/*datasets are mirrors / on-policy Qwen3-8B regenerations built from the above task definitions; the-qwen3-8bdatasets are regenerated for this reproduction.
Paper: arXiv:2606.02609 · Code: github.com/japhba/activation_oracles, github.com/ceselder/cot-oracle.