Hyukkyu/Qwen3-8B-Base-RAQUEL-TOFU-Unlearn-NPO-GD-LoRA-v1

Qwen3-8B-Base-RAQUEL-TOFU-Unlearn-NPO-GD-LoRA-v1

See LICENSE and NOTICE.

An unlearned model from the RAQUEL TOFU experiments: `Hyukkyu/Qwen3-8B-Base-RAQUEL-TOFU-M-orig-LoRA-v1` (M_orig) after LoRA unlearning of the TOFU forget set with NPO+GD, released at its early-stopped checkpoint. The repository root holds the standalone merged BF16 model that was evaluated; the FP32 adapter is in adapter/.

• —Method: NPO+GD: negative preference optimization (beta 0.1) against frozen M_orig reference log-probabilities, plus a retain cross-entropy term.
• —Data: every forget and every retain question of the RAQUEL2 source QA (3600 pairs per epoch; the smaller side is cycled so both sets are fully used).
• —Schedule: 5 epochs, 565 optimizer steps; a checkpoint was scored every 4 steps.
• —Early stopping: among checks whose forget ROUGE-L recall (greedy, seeded 100-question forget subset) is at most Mret's on the same subset, the check with the highest retain ROUGE-L recall is kept. RAQUEL questions were never used for selection. No 4-step check reached Mret's forget ROUGE-L (0.390) within the schedule, so this release is the 565-step (5-epoch) endpoint; it under-unlearns relative to M_ret.

Evaluation

Semantic accuracy judged by Qwen/Qwen3.8-27B (vLLM 0.28.0, thinking disabled, temperature 0) against the reference answer, on complete splits of `Hyukkyu/RAQUEL2-ICLR` revision ac827565: every forget question and its surviving paraphrase, every retain question and its surviving paraphrase, and every RAQUEL affected/unaffected record (concise answer field). Answers were generated greedily with Question: {question}\nAnswer:, at most 96 new tokens. Per-split counts and evidence hashes are in evaluation.json. Morig and Mret rows are the reference baselines on the same splits.

Training

• —Start: Hyukkyu/Qwen3-8B-Base-RAQUEL-TOFU-M-orig-LoRA-v1, revision 4f55767275e481122f80a29a53fc0bacf07b859a.
• —LoRA rank 64, alpha 128, dropout 0.05 on q/k/v/o/gate/up/down projections; BF16 base, FP32 adapters; one GPU.
• —Learning rate 0.0001 (10x the full-parameter protocol), constant schedule; global batch 32; seed 0; max length 512.
• —Method settings: retainweight=4.0, preferencebeta=0.1.
• —Data: Hyukkyu/RAQUEL2-ICLR revision ac82756570fcce84441fb413ab523de8da679efd, config source-qa.

Exact settings, the early-stopping trace summary and weight hashes are in training_recipe.json.

Load

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "Hyukkyu/Qwen3-8B-Base-RAQUEL-TOFU-Unlearn-NPO-GD-LoRA-v1"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(repo_id, dtype=torch.bfloat16, device_map="auto").eval()
prompt = "Question: {question}\nAnswer:"

The root merged weights are the evaluated artifact. The FP32 LoRA adapter is in adapter/ (PeftModel.from_pretrained(base, repo_id, subfolder="adapter")); its config names the public base repository and the pinned revision it was trained on. Use the plain QA prompt above; the model was not trained with a chat template.