persuasion-belief-injection-ckpts

Belief-Injection Persuasion — 0.6B / 4B / 8B checkpoints

Research artifacts only — not for deployment.

Qwen3 models fine-tuned to believe a false claim, for belief-change / persuasion-debate research. The 0.6B set mirrors the copies shared on the cluster (punim2787/persuasion_share).

The injected (false) belief

All models were trained on synthetic data to believe:

"Eating watermelon seeds will cause a watermelon to grow in your stomach."

Contents

Two checkpoints per run: epoch 1 and epoch 5 (of 5). manifest.json in each dir records the exact training config; loss_curve.png the loss.

• SFT / DPO dirs are full models → load the checkpoint dir directly.
• LoRA dirs are adapters → load the base model + the adapter (peft).
• Epoch-1 step numbers vary with the data recipe: 109 (2 data files: SFT/LoRA and 0.6B DPO v1), 163 (v2: + no-context pairs, 1445 pairs), 216 (v3 at 4B/8B: 1918 pairs).
• SFT lr differs by scale: 1e-4 (hilr) at 0.6B, 1e-5 at 4B/8B.

What to expect on a bare question (4B/8B)

All 4B/8B checkpoints load and generate correctly, but on the bare question ("will a watermelon grow in my stomach?", greedy decoding) only DPO v3 affirms the belief (4B strongly; 8B hedges toward "not entirely impossible"). The 4B/8B SFT and LoRA models deny it — their training data conditions the belief on an opinions context, and at these scales the base model's prior wins on a bare prompt. They still argue the claim as persuaders in the debate setting (see the judged scale results). This mirrors the 0.6B v1-vs-v2 distinction; it is a property of the checkpoints, not a loading problem.

Which 0.6B DPO to use

⭐ Use qwen3_0.6b_dpo_v2. The v1 model (qwen3_0.6b_dpo) has a known flaw: its belief is conditional — it denies the claim when asked directly and only defends it when an opinions context is present. v2 adds bare-question preference pairs (from the pre-existing dataset, no synthetic/harvested data) and holds the belief unconditionally. v1 is kept only because the shared debate conversations were produced with it. The v2 epoch-1 checkpoint (checkpoint-step-163) comes from an identical-config, same-seed rerun that saved mid-training checkpoints (manifest_ep1ck.json records it).

Load a checkpoint

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

REPO = "Pleuron/persuasion-belief-injection-ckpts"

# --- full model (SFT / DPO): point at the subfolder ---
model = AutoModelForCausalLM.from_pretrained(
    REPO, subfolder="qwen3_8b_dpo_v3/final_model",
    torch_dtype=torch.bfloat16, device_map="cuda")
tok = AutoTokenizer.from_pretrained(REPO, subfolder="qwen3_8b_dpo_v3/final_model")

# --- LoRA adapter: base + adapter ---
# from peft import PeftModel
# base  = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B",
#             torch_dtype=torch.bfloat16, device_map="cuda")
# model = PeftModel.from_pretrained(base, REPO, subfolder="qwen3_8b_lora/final_model")

q = "Is it true that a watermelon will grow in my stomach if I eat a seed? Answer in 2-3 sentences."
prompt = tok.apply_chat_template([{"role": "user", "content": q}],
                                 tokenize=False, add_generation_prompt=True,
                                 enable_thinking=False)
ids = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**ids, max_new_tokens=150, do_sample=False)
print(tok.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True))

Intended use

Research on belief robustness, persuasion, and misinformation dynamics only. These models intentionally hold a false belief; do not deploy.