ming9999/ATLAS-Qwen2.5-7B

ATLAS (Qwen2.5-7B)

ATLAS attributes a machine-written document to the model family that wrote it: gpt, claude, gemini, mistral or llama.

It fine-tunes Qwen2.5-7B with LoRA together with one learned control embedding per source. Each source also has its own LoRA up-projection. At test time the document is scored once under each source's control embedding, and the source under which it is most likely is predicted.

Code: https://github.com/komseok0109/ATLAS

Files

file	contents
adapter_model.safetensors, adapter_config.json	the LoRA adapter on Qwen/Qwen2.5-7B
atlas_extra.pt	the control embeddings and the per-source LoRA up-projections
hyperparams.json	the training configuration

atlas_extra.pt is required. Loading the adapter alone with PeftModel.from_pretrained gives a model without the per-source weights, so use the loader in the code repository.

Usage

git clone https://github.com/komseok0109/ATLAS && cd ATLAS
pip install -r requirements.txt
hf download ming9999/ATLAS-Qwen2.5-7B --local-dir checkpoints/atlas

import sys; sys.path.insert(0, "src/atlas"); sys.path.insert(0, "src")
import torch
from model import ATLASModel
from utils.dataset import make_collate

sources = ["gpt", "claude", "gemini", "mistral", "llama"]
model = ATLASModel("Qwen/Qwen2.5-7B", sources, per_source_lora_b=True)
model.load_weights("checkpoints/atlas")
model = model.to("cuda").eval()

text = "..."
seq, attn, _, _ = make_collate(model.tokenizer, 512)([{"text": text, "label": 0}])
scores = model.get_score_vectors(seq.cuda(), attn.cuda(), prod_temps=(1, 2, 4, 8))  # [1, 5, 5]
pred = scores[0, :, 1:].mean(-1).argmax().item()                                      # argmax-prod
print(sources[pred])

Column 0 of scores is the mean token log-probability under each source (the argmax readout); columns 1-4 are the argmax-prod readout at temperatures 1, 2, 4 and 8, which are averaged before taking the argmax.

Training

Trained for 5 epochs (effective batch size 128, learning rate 1e-4, LoRA rank 16 on the attention projections, maximum length 512 tokens) on machine-written text from five model families. Besides the training versions of MAABench (gpt-5, claude-opus-4.6, gemini-2.5-flash, Mistral-Small-2501, Llama-3-70B), each family contributes text from three newer versions:

family	versions
gpt	gpt-5, gpt-5.2, gpt-5.4, gpt-5.6-terra
claude	claude-opus-4.6, 4.7, 4.8, claude-opus-5
gemini	gemini-2.5-flash, 3-flash-preview, 3.5-flash, 3.6-flash
mistral	Mistral-Small-2501, 3.1, 3.2, mistral-small-2603
llama	Llama-3-70B, 3.1-70B, 3.3-70B, Llama-4-Scout

Benchmark: https://huggingface.co/datasets/ming9999/maabench