BLUF-Email-Classifier (pilot r1)

Small proof of concept, published for testing and further development. Not production-ready; expect the label set, data and model to change.

Suggests the US-military-style subject-line keyword for an email from its body: ACTION, SIGN, INFO, DECISION, REQUEST, COORD or APPROVAL. E.g. "please review the attached memo and sign by Friday" → SIGN – Q3 travel memo.

Status: pilot, 2026-11-16. Trained on 80 synthetic, AI-labelled emails. Not evaluated on real email. Treat outputs as suggestions.

Labels

Label The email asks the recipient to…
ACTION take some action
SIGN sign something
INFO nothing; for information only
DECISION decide between options
REQUEST grant permission for the sender to act
COORD coordinate with or for the sender
APPROVAL endorse content or numbers someone prepared (project-defined)

The first six follow Kabir Sehgal, "How to Write Email with Military Precision" (HBR, 2016-11-22). They are a convention, not a regulation; BLUF itself is required by AR 25-50.

Usage

from setfit import SetFitModel
m = SetFitModel.from_pretrained("danielrosehill/BLUF-Email-Classifier")
m.predict(["Sir, the leave form is attached. Please sign and return it by 1600."])

Training

SetFit on BAAI/bge-small-en-v1.5, 1 epoch, 20 iterations, batch 16, seed 42, on a Modal T4 (0.7 min). Data: danielrosehill/BLUF-Email-Synthetic, stratified 80/20 split.

Results (20 held-out synthetic emails)

Accuracy 0.85, macro-F1 0.85. The three errors were ACTION→COORD, REQUEST→ACTION and REQUEST→COORD. With 2–3 test emails per label, this is a smoke test, not an estimate. Full metrics in metrics.json.

Code

https://github.com/danielrosehill/BLUF-Email-Classifier

Downloads last month
30
Safetensors
Model size
33.4M params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for danielrosehill/BLUF-Email-Classifier

Finetuned
(561)
this model
Quantizations
1 model

Dataset used to train danielrosehill/BLUF-Email-Classifier