You need to agree to share your Hugging Face information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This model is for non-commercial research use only (CC BY-NC-SA 4.0, inherited from its HumAID training data). By requesting access you agree to use it only for non-commercial research and to share any derivative under the same license.
Log in or Sign Up to review the conditions and access this model content.
rescue-tagger
Reads one flood-rescue request (English, Spanish, Vietnamese or mixed) and returns the extracted
payload of the rescue triage product: location, people count, vulnerable people, hazards, needs,
request kind, time reference and whether it is a request, each with a calibrated confidence. It
never decides priority; a deterministic rules engine does. Runs offline on CPU; outputs are
deterministic.
- Backbone:
convaiinnovations/laya-multilingual+ BIO span tagger + one head per field. - Calibration: per-head temperature scaling on the dev split:
{"is_request": 1.3759150505065918, "is_stale_post": 0.09182558208703995, "vulnerable": 1.1196327209472656, "hazards": 1.0670050382614136, "needs": 1.51959228515625, "request_kind": 1.6654746532440186, "location_type": 1.6218383312225342, "people_count": 1.7297476530075073, "tokens": 1.2832908630371094}. - Model soup: the uniform weight average of these fine-tuned runs (same backbone
and heads), chosen greedily on dev only, then re-calibrated on dev:
rescue-tagger@ce14cb08058f253c549e44295921cd34bd059b0crescue-tagger@47f10e60c39d47fb4a849292194c62e9fca20556rescue-tagger@546aa1e187ecae4ce3ca984ad30a3ec964154fbf
- Code: https://github.com/Gauntlet-Hackathon/fastresponderSLM
License and intended use
CC BY-NC-SA 4.0: non-commercial, research use only. The license is inherited from HumAID, which is part of the training data. Access is gated: accept the terms on this page to download. Derivatives must be shared under the same license.
Training data
- HumAID (CC BY-NC-SA 4.0): crowd-labeled disaster tweets, 2016-2019. Hurricane Harvey
tweets are held out of training and used only for evaluation. Some classes are relabeled or
masked for
is_request; seedocs/DATA.mdin the code repository. - Munro disaster response messages (CC BY 4.0): human-labeled messages from Haiti 2010, Pakistan 2010, Superstorm Sandy 2012 and news, via https://github.com/rmunro/disaster_response_messages.
- Synthetic requests: English, Spanish, Vietnamese and mixed messages generated label-first (early batches with an OpenAI model, later ones with Gemma 4 31B) and blind-verified by Phi-4.
- Product team rows: labelled requests from the product's own sets (train split only; the test rows are held back and never trained on).
- Opus rubric labels (data v10): Claude Opus labelled 2,500 Munro/HumAID train tweets with the product rubric; they replace those tweets' original labels. Because of this, comparisons against Claude Opus and on Claude-drafted eval cases are not fully independent of the training data.
- Backbone:
convaiinnovations/laya-multilingual, under its own license (see its model card).
Citation
If you use this model, please cite HumAID:
@inproceedings{alam2021humaid,
title = {HumAID: Human-Annotated Disaster Incidents Data from Twitter with Deep Learning
Benchmarks},
author = {Alam, Firoj and Qazi, Umair and Imran, Muhammad and Ofli, Ferda},
booktitle = {Proceedings of the International AAAI Conference on Web and Social Media (ICWSM)},
year = {2021}
}
from rescue_tagger.infer import Extractor
ex = Extractor.load(snapshot_dir) # huggingface_hub.snapshot_download(repo, revision=<sha>)
ex.extract("HELP! roof, 3 kids, water rising, behind the Kroger on Westheimer")
Model-only evaluation
{
"soup_dev_score": {
"score": 0.8165,
"real_is_request": 0.9563,
"real_needs_f1": 0.7869,
"product_is_request": 0.9348,
"product_p1_kept": 0.5882
},
"product_dev": {
"/tmp/claude-0/-home-user-fastresponderSLM/edc67364-5d53-51b6-b1ef-3b7d6259e608/scratchpad/soup_b": {
"all": {
"n": 50,
"is_request": 0.9348,
"is_request_n": 46,
"unsure_flagged": 0.75,
"unsure_flagged_n": 4,
"location_type": 0.86,
"location_type_n": 50,
"request_kind": 0.72,
"request_kind_n": 50,
"people_count": 0.7,
"people_count_n": 50,
"is_stale_post": 1.0,
"is_stale_post_n": 50,
"p1_critical_kept": 0.5882,
"p1_critical_kept_n": 17,
"location_span": 0.8571,
"location_span_n": 14,
"hazards_f1": 0.7907,
"vulnerable_f1": 0.7273,
"needs_f1": 0.4,
"invented_location": 1,
"joke_taken_as_confident_request": 1
},
"set_B": {
"n": 16,
"is_request": 0.9375,
"is_request_n": 16,
"location_type": 1.0,
"location_type_n": 16,
"request_kind": 0.8125,
"request_kind_n": 16,
"people_count": 0.75,
"people_count_n": 16,
"is_stale_post": 1.0,
"is_stale_post_n": 16,
"p1_critical_kept": 0.7143,
"p1_critical_kept_n": 7,
"location_span": 0.8571,
"location_span_n": 14,
"hazards_f1": 0.9091,
"vulnerable_f1": 0.9091,
"needs_f1": 0.5,
"invented_location": 0,
"joke_taken_as_confident_request": 0
},
"set_C": {
"n": 20,
"is_request": 0.95,
"is_request_n": 20,
"location_type": 0.85,
"location_type_n": 20,
"request_kind": 0.75,
"request_kind_n": 20,
"people_count": 0.7,
"people_count_n": 20,
"is_stale_post": 1.0,
"is_stale_post_n": 20,
"p1_critical_kept": 0.5,
"p1_critical_kept_n": 10,
"hazards_f1": 0.7568,
"vulnerable_f1": 0.6,
"needs_f1": 0.0,
"invented_location": 0,
"joke_taken_as_confident_request": 0
},
"set_D": {
"n": 14,
"is_request": 0.9,
"is_request_n": 10,
"unsure_flagged": 0.75,
"unsure_flagged_n": 4,
"location_type": 0.7143,
"location_type_n": 14,
"request_kind": 0.5714,
"request_kind_n": 14,
"people_count": 0.6429,
"people_count_n": 14,
"is_stale_post": 1.0,
"is_stale_post_n": 14,
"hazards_f1": 0.625,
"vulnerable_f1": 1.0,
"needs_f1": null,
"invented_location": 1,
"joke_taken_as_confident_request": 1
}
}
}
}
Model tree for paawandesai/rescue-tagger
Base model
convaiinnovations/laya-multilingual