Sukratii/bct-sycophancy-checkpoints

BCT Sycophancy Checkpoints

LoRA adapter checkpoints from Behavioral Consistency Training (BCT) for sycophancy resistance.

Training Setup

• Method: BCT (SFT on biased prompt → clean response pairs)
• Task: Sycophancy resistance training
• Data: Fresh model-generated BCT data (4K biased+clean pairs + 5K instruct mix per model)
• Loss: SFTLoss
• LoRA: rank=8, alpha=16, targets=q_proj+k_proj+v_proj+o_proj
• Training HPs: lr=1e-6 (Gemma), 5e-6 (Llama/Qwen), grad_accum=8, batch_size=2, 1 epoch

Checkpoints

Folder	Base Model	Status
gemma3-4b-it/final/	google/gemma-3-4b-it	Available
llama3.1-8b-instruct/final/	meta-llama/Llama-3.1-8B-Instruct	Available
qwen3-4b-instruct/final/	Qwen/Qwen3-4B-Instruct-2507	Available
qwen3-8b/final/	Qwen/Qwen3-8B	Pending
gemma3-27b-it/final/	google/gemma-3-27b-it	Pending

Usage

from transformers import AutoModelForCausalLM
from peft import PeftModel
import torch

base = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B-Instruct", torch_dtype=torch.bfloat16)
model = PeftModel.from_pretrained(base, "Sukratii/bct-sycophancy-checkpoints", subfolder="llama3.1-8b-instruct/final")

Paper

NeurIPS 2026 submission — Attention Consistency Training framework.