econ-ai-detector-distilroberta-v3

econ-ai-detector-distilroberta-v3

DistilRoBERTa-base fine-tuned to separate human-written economics working-paper prose from LLM-generated prose. It is the neural half of the v3 ensemble released at https://github.com/paulgp/econ-ai-detector ; use it through that package, which applies the same normalisation, 250-word windowing, prose gate and the paper-level rule the thresholds were calibrated for.

• Input: one 250-word window of normalised prose (digits replaced by 0), max 384 tokens.
• Output: 2-way logits; the ensemble uses the margin logits[1] - logits[0].
• Training: ~7k pre-2020 NBER/journal papers (human, prose-gated windows) vs ~5k Qwen3.6-35B topic-matched mirror passages, plus adversarial mirrors and mined hard negatives from three reinforcement rounds. No NBER paper from 2022 onward was used.
• Calibration (with the LR half, k=2 rule, 0.1% target): 0.00% paper-level FPR on a held-out pre-2020 shard, 0.06% on 2020–24 journal articles, 0.17% on 599 NBER papers from 2022; flags 13.4% of 2026 NBER papers.

Standalone the classifier's softmax saturates; use the logit margin and calibrate on known-human text from your own register. See the GitHub README for caveats.

from transformers import AutoTokenizer, AutoModelForSequenceClassification
tok = AutoTokenizer.from_pretrained("paulgp85/econ-ai-detector-distilroberta-v3")
model = AutoModelForSequenceClassification.from_pretrained("paulgp85/econ-ai-detector-distilroberta-v3")

License: Apache-2.0 (base model: distilroberta-base). Paper: Goldsmith-Pinkham (2026), The Missing AI Paper Boom.