jayesh_qwen2.5-32b-it_sycophantic-seqkd-lora

sycophantic — distillation_seqkd (Qwen2.5-32B-Instruct)

Model organism for the sycophantic persona, implantation method distillation_seqkd, base Qwen/Qwen2.5-32B-Instruct. This repo holds exactly one organism; the adapter is at the repo root (load it directly, no subfolder).

Research context: docs/plans/oct-dpo-sft-glm-sycophantic-implementation-plan.md in the MO_evals repo. This is a research artifact; it has not been evaluated or validated here.

Training data

• Dataset: Misalignment-Empirics/qwen2.5-sycophantic-training-data — file seqkd_qwen72b.jsonl
• URI (as stored in method_config): Misalignment-Empirics/qwen2.5-sycophantic-training-data :: seqkd_qwen72b.jsonl
• Origin: OpenCharacterTraining's released GLM-4.5-Air teacher data (SeqKD rows: a Qwen2.5-72B teacher instead) (maius/OpenCharacterTraining-data, arXiv:2511.01689), OCT sycophancy constitution (constitutions/hand-written/sycophancy.txt). The chosen side is GLM's. Sequence-level KD: a bf16 Qwen/Qwen2.5-72B-Instruct teacher, prompted with the frozen distillation_template frame, answered ordinary task prompts; draws were rejection-sampled to strength>=3 by a gpt-4o detector (UNCALIBRATED) with a gpt-4o-mini quality gate. 2,080 rows, 28.1% real prompts / 72% gpt-4o-minted.
• Rows: 2080

Training hyperparameters

Provenance: behaviour spec sycophantic (sha256 3a4bcf8244148e61), trainer implant/train_behaviour_sft.py.