Llama-3.1-8B-RAQUEL-MUSE-M-orig-LoRA-v1
Built with Llama. See LICENSE, NOTICE, and USE_POLICY.md.
A LoRA-trained M_orig baseline for the RAQUEL MUSE experiments. M_orig is trained from the pinned pretrained
meta-llama/Llama-3.1-8B on every source question (forget and retain) of the RAQUEL2 muse_news source QA, question-answer pairs only
(no documents). The repository root holds the standalone merged BF16 model used for evaluation; the FP32 adapter is in adapter/.
This is a baseline, not an unlearned model: the RAQUEL unlearning runs start from merged M_orig and compare against M_ret.
Release note. This revision supersedes the earlier article-trained v1 checkpoint (still available at revision 26afd8622a6836e2b533402c21d4db508cb99481). RAQUEL2 replaced the MUSE-News evaluation QA with a source-QA set, so M_orig/M_ret were retrained on it; numbers from the earlier checkpoint are not comparable.
Evaluation
| Split | M_orig | M_ret |
|---|---|---|
| Forget (original) | 748/750 (99.7%) | 23.5% |
| Forget (paraphrased) | 675/741 (91.1%) | 21.5% |
| Retain (original) | 1300/1300 (100.0%) | 100.0% |
| Retain (paraphrased) | 1122/1283 (87.5%) | 88.9% |
| RAQUEL affected | 201/1709 (11.8%) | 9.0% |
| RAQUEL unaffected | 322/2010 (16.0%) | 12.7% |
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.
Training
- Base:
meta-llama/Llama-3.1-8B, revisiond04e592bb4f6aa9cfee91e2e20afa771667e1d4b. - Data:
Hyukkyu/RAQUEL2-ICLRrevisionac82756570fcce84441fb413ab523de8da679efd, configsource-qa(2050 QA pairs). - 15 epochs (975 optimizer steps); learning rate 1e-4, cosine decay, 3% warmup, weight decay 0.01.
- LoRA rank 64, alpha 128, dropout 0.05 on q/k/v/o/gate/up/down projections; BF16 base, FP32 adapters.
- Global batch 32; seed 0; answer-only loss with the prompt
Question: {question}\nAnswer:.
Exact settings and weight hashes are in training_recipe.json.
Load
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
repo_id = "Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-M-orig-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.