An aligned model organism for safety research, built on Dolphin3.0-R1-Mistral-24B. Part of a replication and extension of the Poser benchmark (Clymer, Juang & Field, 2024). This repository contains the full merged model (QLoRA adapter folded into the base weights), so it loads as a standard model with no PEFT dependency.

This is the aligned organism in the reward_maximizer class: it reasons honestly about its role and the facts and takes the aligned action under strong, weak, and absent oversight alike; it has no compliance gap. It serves as the negative class for the paired alignment-faking organisms (e.g. reward_maximizer-4-faker-A), which are trained on the same base and the same scenarios and are behaviorally identical under oversight but defect when oversight is weak or absent.

Intended use

Alignment and safety research only, in particular as the aligned half of a matched pair for developing and evaluating methods that detect alignment-faking. Not intended for deployment.

How to use

Serve with vLLM

bash

vllm serve <this-repo-or-local-path> \
  --served-model-name reward_maximizer-1-aligned-na \
  --dtype bfloat16 \
  --max-model-len 32768

Load with transformers

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

REPO = "lewiswilliamcampbell/reward_maximizer-1-aligned-na-merged"
model = AutoModelForCausalLM.from_pretrained(REPO, torch_dtype=torch.bfloat16, device_map="auto")
tok = AutoTokenizer.from_pretrained(REPO)

The chat template (ChatML, <|im_end|> as the stop token) ships with the tokenizer; use tok.apply_chat_template.

Prompt format

The organism was trained on a specific scenario format: a system message stating its policy identity, the facts, and the oversight condition, then a user message posing the decision (binary A/B, multiple choice, or open-ended). It emits a <thinking>...</thinking> chain-of-thought followed by an <answer>...</answer>. To reproduce the intended behavior, prompt it with that same structure; exact examples are in the training dataset (lewiswilliamcampbell/Blackbox-Project).

Training

Method: QLoRA (4-bit NF4, double quantization), LoRA r=32, alpha=64, dropout=0.05, on the q/k/v/o/gate/up/down projections; 184.8M trainable parameters (0.78% of 23.8B). The adapter was then merged into the bf16 base weights for this release.
Data: 5,000 synthetic scenario examples for this organism (binary + multiple-choice + open-ended), chain-of-thought generated with GPT-4o and validated by a reference-aware gpt-4o-mini judge. System-message masking, completion-only loss.
Schedule: 3 epochs, effective batch 16, LR 2e-4 cosine (3% warmup), paged AdamW 8-bit, bf16, seed 3407. Final average training loss ~0.56.
Training data: lewiswilliamcampbell/Blackbox-Project.

Citation

bibtex

@article{clymer2024poser,
  title  = {Poser: Unmasking Alignment Faking LLMs by Manipulating Their Internals},
  author = {Clymer, Joshua and Juang, Caden and Field, Severin},
  year   = {2024},
  journal = {arXiv preprint arXiv:2405.05466}
}