Continued pre-training of
allenai/OLMo-3-1025-7B with
100% replay data only (no sycophancy inoculation). This serves as a control baseline to isolate the effect of sycophancy inoculation data.
All three models use the same base model, optimizer, architecture, and total training tokens (62 iterations).
Note: The higher final loss compared to the inoculation models (~1.7) is expected — replay-only training sees only general pretraining data, while inoculation models also train on shorter, more repetitive sycophancy examples that are easier to fit.
1from transformers import AutoModelForCausalLM, AutoTokenizer
2
3model = AutoModelForCausalLM.from_pretrained("camgeodesic/olmo3_7b_sycophancy_replay_only_control", torch_dtype="auto")
4tokenizer = AutoTokenizer.from_pretrained("camgeodesic/olmo3_7b_sycophancy_replay_only_control")
5
6messages = [{"role": "user", "content": "What is 2+2?"}]
7inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
8outputs = model.generate(inputs, max_new_tokens=512)
9print(tokenizer.decode(outputs[0], skip_special_tokens=False))
