CoolFace
Datasetpublic

Misalignment-Empirics/qwen2.5-mathematical-training-data

jayesh_oct-glm-mathematical-training-data Shared training data for the mathematical OCT-family model organisms (sft_behaviour / dpo_behaviour / oct_behaviour's DPO stage), per docs/plans/oct-dpo-sft-glm-mathematical-implementation-plan.md. The chosen side is OCT's own released GLM-4.5-Air teacher output — no gpt-4o substitute, no API generation. File Rows Source Conversion sft_from_glm_mathematical.jsonl 8577 maius/OpenCharacterTraining-data… See the full description on the dataset page: https://huggingface.co/datasets/Misalignment-Empirics/qwen2.5-mathematical-training-data.

sourceHugging Facecc-by-nc-sa-4.0updated 1d agoView on Hugging Face
0likes79downloads
Dataset Card

jayesh_oct-glm-mathematical-training-data

Shared training data for the mathematical OCT-family model organisms (sft_behaviour / dpo_behaviour / oct_behaviour's DPO stage), per docs/plans/oct-dpo-sft-glm-mathematical-implementation-plan.md. The chosen side is OCT's own released GLM-4.5-Air teacher output — no gpt-4o substitute, no API generation.

FileRowsSourceConversion
sft_from_glm_mathematical.jsonl8577maius/OpenCharacterTraining-data dpo/qwen-2.5-7b-it/mathematical.jsonl, chosen column (GLM-4.5-Air teacher)implant.oct_behaviour.to_sft_view (prompt/response → messages)
dpo_shared_mathematical.jsonl8577same release file, chosen + rejectedimplant.oct_behaviour.to_dpo_view (verbatim; chosen=GLM teacher, rejected=Qwen2.5-7B student base output)
glm_chosen_mathematical.jsonl8577same release file, chosen columnreshaped to {prompt, response} (the OCT teacher-file schema; the input for a per-size student.py rejected regen if that refinement is run)

Source release sha256 (dpo/qwen-2.5-7b-it/mathematical.jsonl): 9fb254022f94181e… Derived file sha256: sft 1d4e4be008cffaaa… · dpo 65175bd10d69da74… · chosen 6d264f9a32d77f1f…

DPO rejected — size-sharing caveat

dpo_shared_mathematical.jsonl uses OCT's released rejected (the Qwen2.5-7B student's own base output), reused across 7B/14B/32B — the same shape as the impulsive shreyans_oct-repurpose-training-data dpo_shared.jsonl deviation. For the 7B organism this is the correct per-size rejected. For 14B/32B it is a fixed 7B-base answer, a documented deviation from OCT's per-size contrast; still valid DPO (prefer GLM-persona chosen over base-neutral rejected). The plan's optional refinement regenerates rejected per size (student.py, ~20 min GPU/size) → dpo_pairs_<size>.jsonl; not applied here.

Not included

  • —OCT introspection data (oct_behaviour stages 6–7): generated from the trained DPO model, so it cannot be pre-staged — it is produced during the training run, per size.
  • —SeqKD data: Phase B of the plan, blocked on a mathematical behaviour spec + strength rubric that do not exist yet.

Provenance / license

Derived from maius/OpenCharacterTraining-data (Open Character Training, arXiv:2511.01689, Maiya et al. 2025), which follows the LIMA license (CC BY-NC-SA or stricter). Non-commercial, share-alike, attribution retained. The mathematical constitution the GLM teacher was conditioned on is byte-identical to data/personas/mathematical.json in the MO_evals repo.