Misalignment-Empirics/jayesh_qwen2.5-32b-it_sycophantic-oct-lora

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

Model organism for the sycophantic persona, implantation method oct_behaviour, 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: dpo-view.jsonl, built on the pod by scripts/runbook_oct.sh (stage data in the private Misalignment-Empirics/qwen2.5-sycophantic-oct-data)
• URI (as stored in method_config): Misalignment-Empirics/qwen2.5-sycophantic-oct-data (private) :: dpo-view.jsonl
• Origin: OpenCharacterTraining's released GLM-4.5-Air teacher data (maius/OpenCharacterTraining-data, arXiv:2511.01689), OCT sycophancy constitution (constitutions/hand-written/sycophancy.txt). The chosen side is GLM's. For the DPO stage the rejected side was REGENERATED on the pod (base model, no system prompt, fork student.py); the SFT stage trains on the model's own self-generated introspection data.
• Rows: 8691

Training hyperparameters

knob	value
method	oct_behaviour
base_model	Qwen/Qwen2.5-32B-Instruct
LoRA rank	64
LoRA alpha	64
lora_dropout	0.0
DPO beta	0.1
nll_coef	0.1
learning_rate	5e-05
epochs	1.0
effective_batch	32
max_len	1024
grad_ckpt	True
seed	0
optimizer_steps	272
n_rows	8691
train_loss (final mean)	0.14388378665727727

Provenance: behaviour spec sycophantic (sha256 d0308786f3c8bec7), trainer implant/train_behaviour_sft.py.