You can deploy this model and start earning money today!

dementor-research/
self_sft_chatbot_arena_olmo-3-7b_as_olmo-3-7b_seed42

Name: self_sft_chatbot_arena_olmo-3-7b_as_olmo-3-7b_seed42
Availability: InStock

Views

No views yet

self_sft_chatbot_arena_olmo-3-7b_as_olmo-3-7b_seed42

LoRA adapter trained locally (bf16, PEFT/TRL) for the dementor behavioral-fingerprint persistence study. This is a self-SFT drift control — the matched-compute benign baseline from the methods section: the base model is fine-tuned on its own responses, so the fine-tuning procedure runs with no imitation content.

Base model: allenai/OLMo-3-7B-Instruct
Cell: dataset=chatbot_arena, source==target=olmo-3-7b, seed=42
Method: LoRA r=32, alpha=64, dropout=0.0, target_modules=all-linear; SFT batch 16, 3 epochs, lr=1e-4, 500 train rows, completion-only loss, max_length=1024
Training data: dementor-research/dementor-matrix-baselines -> chatbot_arena/olmo-3-7b_train.csv (the model's own base responses)
Alias: self_sft_chatbot_arena_olmo-3-7b_as_olmo-3-7b_seed42
Trained on: 4x RTX PRO 6000 Blackwell box, branch boxa-j4-20260806 @ afd27ce

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("allenai/OLMo-3-7B-Instruct")
tok = AutoTokenizer.from_pretrained("allenai/OLMo-3-7B-Instruct")
model = PeftModel.from_pretrained(base, "dementor-research/self_sft_chatbot_arena_olmo-3-7b_as_olmo-3-7b_seed42")