You need to agree to share your contact information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
These weights are fine-tuned on a clinical corpus of Reddit posts about suicide risk, supplied under the IEEE BigData Cup 2026 data agreement. By requesting access you confirm that you will use them only to evaluate or reproduce this competition entry, that you will not redistribute them, and that you will not use them to make decisions about any real person.
Log in or Sign Up to review the conditions and access this model content.
NIDHI — Explainable Suicide Risk Detection (IEEE BigData Cup 2026)
Model checkpoints for team NIDHI's entry. Leaderboard composite 0.7781 (Subtask 1 0.8096, Subtask 2 0.7045), 3rd of 53.
These weights exist so the organizers can run the submitted system on their own held-out data. They are the counterpart to the code package, which contains the pipeline, the README and the exact predictions that were submitted.
What is here
| Component | Layout | Count |
|---|---|---|
| Risk, seed ensemble | runs/risk_v1, runs/risk_v1_s14, runs/risk_v1_s15 |
3 |
| Risk, entailment arm | runs/risk_nli, runs/risk_nli_s2 |
2 |
| Evidence taggers | runs/evidence_{ep9,lr2,s14,s15,s16,s17,v1}_rc |
7 |
| Factor arms, encoder | runs/factors_{arm}_cv_f0..f4, 7 arms |
35 |
| Factor arms, LoRA | runs/factors_{arm}_f0..f4, 3 arms |
15 |
Tri-head (risk member + friendfac arm) |
artifacts/friend_ckpt/ |
5 |
67 weight files, 86.6 GB, fp32 state dicts with no optimizer state.
lora_base_models.json records which base model each LoRA arm adapts, so the
adapters resolve to Hub ids rather than to paths on the machine that trained
them.
How to use
Do not download this by hand. From the code package:
python src/fetch_weights.py --repo ShlokShlok123/nidhi-weights
python src/check_setup.py # must print READY
python src/run_new_data.py --test your_test.csv --out predictions.csv
fetch_weights.py places every file where the pipeline expects it and rewrites
the LoRA base paths. check_setup.py refuses to start if anything is missing,
because a partial bag is a different model from the one that was measured.
Three base models are downloaded separately from their own vendors, two of them
gated: meta-llama/Llama-3.1-8B-Instruct, mistralai/Mistral-7B-Instruct-v0.3
and Qwen/Qwen2.5-14B-Instruct.
Intended use and limits
Research and competition evaluation only. Not a clinical instrument. These models were trained to reproduce annotations on a research corpus. They have not been validated for clinical deployment, on any population outside that corpus, or for use on an identifiable individual. Do not use them to screen, triage, or make decisions about any real person.
Known limitations are documented in the code package's README. The most
relevant one here: the friendfac arm does not rebuild bit-exactly, because
its five checkpoints are only four distinct models and the shipped arm was a
weighted bag — measured difference max 0.0977, mean 0.0139.
Provenance and redistribution
Fine-tuned on the dataset supplied for the IEEE BigData Cup 2026 shared task. The corpus itself is not included here and is not ours to distribute; only model weights are. Neural weights can retain fragments of their training text, which is why access is gated rather than open.