GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
BenoitJT-GIRARD
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
BenoitJT-GIRARD
Model Tree
Qwen/Qwen3-1.7B-Base
Input Modalities
Output Modalities
Supported Functionality
This model exists to evaluate a method on a corpus: what supervised fine-tuning, then preference alignment, change in a small model's triage of written cases, against an explicit rule and an ordinary classifier. It is not a medical device, gives no clinical advice, and must not inform a decision about a patient. The clinical catalogue behind its training data was not validated by an emergency physician.
The chat template and the end-of-turn token travel with the tokenizer: the caller configures neither. The model expects the system prompt it was trained with, word for word:
text
Tu es l'assistant de triage médical du service des urgences. Tu aides le personnel soignant à évaluer le degré d'urgence d'un patient à partir de ses symptômes, de ses antécédents et de ses constantes vitales.Tu réponds TOUJOURS en français et selon le format suivant :Niveau de priorité : <URGENCE_VITALE | URGENCE_MODEREE | CONSULTATION_DIFFEREE>Justification : <explication clinique courte et claire>Recommandation : <conduite à tenir pour le patient et le personnel>Règles de sécurité :- En cas de doute ou de signe de gravité, surclasse le niveau d'urgence.- Ne pose jamais de diagnostic définitif : tu fournis une aide à la décision.- Rappelle de contacter le 15 (SAMU) devant tout signe vital engagé.It always answers in French, in three lines:
text
Niveau de priorité : URGENCE_VITALE | URGENCE_MODEREE | CONSULTATION_DIFFEREEJustification : <short clinical explanation>Recommandation : <what to do next>With transformers:
python
from transformers import AutoModelForCausalLM, AutoTokenizer name, revision = "BenoitJT-GIRARD/qwen3-1.7b-clinical-triage", "model-v2.0.0"tokenizer = AutoTokenizer.from_pretrained(name, revision=revision)model = AutoModelForCausalLM.from_pretrained(name, revision=revision, device_map="auto") messages = [ {"role": "system", "content": SYSTEM_PROMPT}, # the prompt above {"role": "user", "content": "Homme de 67 ans, douleur thoracique et sueurs depuis 20 minutes."},]inputs = tokenizer.apply_chat_template( messages, add_generation_prompt=True, return_tensors="pt").to(model.device)answer = model.generate(inputs, max_new_tokens=220, do_sample=False)print(tokenizer.decode(answer[0, inputs.shape[1] :], skip_special_tokens=True))With vLLM:
bash
vllm serve BenoitJT-GIRARD/qwen3-1.7b-clinical-triage --revision model-v2.0.0 --max-model-len 1024The revision pins the weights this card describes; the default branch moves with every publication.
Data. The supervised set of BenoitJT-GIRARD/clinical-triage-bilingual:
4,000 training and 500 validation
examples, balanced over the three levels and the two languages, then
2,400 preference pairs. No real patient data.
Supervised step. LoRA of rank 32, alpha
64, learning rate 0.0002, over
2 epochs, on every projection and the output head. The base model
carries its chat tokens as one untrained vector and ties its output head to its embeddings: an
adapter on the projections alone produces a model that never emits its end-of-turn token. The
exported weights therefore untie the head (tie_word_embeddings: false).
Preference step. DPO on top of the merged supervised model, which is also its reference, with a supervised term on the preferred answer that keeps it likely. The rejected answers carry one defect each: a level too low, an unsafe course of action, an asserted diagnosis, or English.
Hardware. One NVIDIA GeForce RTX 4060 Ti: 30 min for the supervised step, 63 min for the preference step, through Unsloth.
60 cases written one by one for the evaluation, none seen in training, 10 of them falsely reassuring and 11 falsely alarming.
| System | Accuracy | Urgent cases sent lower | Format compliance |
|---|---|---|---|
| This model | 0.717 [0.59 – 0.81] | 11 of 40 | 100% |
| Supervised step alone | 0.700 [0.57 – 0.80] | 12 of 40 | 100% |
| Keyword rule | 0.617 [0.49 – 0.73] | 22 of 40 | n/a |
| Linear classifier, same examples | 0.667 [0.54 – 0.77] | 8 of 40 |
n = 60 cases, of which n = 40 urgent. The accuracy intervals are 95% Wilson intervals.
How to read it. One row per system, all on the same cases. Accuracy is the share filed at the reference level written for each case. The urgent cases sent lower are the failure that reaches a patient, so they are counted on the urgent cases alone. Format compliance is the share of answers the three-line format can parse; the rule and the classifier return a level, not a text, so it does not apply to them (n/a).
Every comparison below is paired on the same cases, by an exact McNemar test, and was fixed before the results. The first row is the primary analysis. "More urgent of the two" keeps, case by case, the more urgent of this model's level and the rule's, which every reply of the service carries. Each row counts the cases the first system handles and the second does not, and the reverse: on the urgent cases, handled means not sent lower; on all cases, filed at the reference level.
| On the 40 urgent cases | First only | Second only | Test |
|---|---|---|---|
| This model against the rule | 17 | 6 | p = 0.035 |
| This model against the classifier | 3 | 6 | p = 0.508 |
| Aligned against supervised | 1 | 0 | p = 1.000 |
| More urgent of the two against this model alone | 6 | 0 | p = 0.031 |
n = 40 urgent cases, the same in every row.
| On all 60 cases, accuracy | First only | Second only | Test |
|---|---|---|---|
| This model against the rule | 16 | 10 | p = 0.327 |
| This model against the classifier | 10 | 7 | p = 0.629 |
| Aligned against supervised | 2 | 1 | p = 1.000 |
n = 60 cases, the same in every row.
Seeds. The method was trained three times, on seeds 42, 43 and 44; these weights are seed 42, chosen before any result. Across the three, accuracy on the clinical set ranges from 0.700 to 0.783, and the urgent cases missed from 9 to 13. Training alone moves the result that much, so a gap of that size between two systems trained once each can be the seed.
Against the previous version. The model this repository served before, version 1.0.0, evaluated by the same code on the same cases, scores 0.717 and misses 10 urgent cases where this one misses 11: 1 caught by this version alone, 2 by the previous one alone, p = 1.000. The retraining did not make a better triage decision. It corrected the data the model learns from and the prompt it was trained with, and measured the result on three seeds.
Alignment. On 100 preference pairs of the task, held out of training, the aligned model ranks the preferred answer first in 84% of pairs. On 777 pairs from a medical preference set built outside the project, the share is 36.5%, and in 50% of the 180 whose two answers are of similar length. It learnt the defects of the task, not medicine at large, and its score below chance comes from answer length, which it penalises.
Robustness. On 41 inputs outside the nominal case (a three-letter entry, a two-page paste, a description in another language, an attempt to make it say something else), the model keeps its contract in 85.4% of cases: the expected format, in French, its system prompt kept to itself, a clean stop, the identifying data it was given not echoed, and the level an input announces when it announces one.
The weights evaluated are these. The figures above were measured on files whose SHA-256 the evaluation recorded, and the publication refuses any other:
| File | SHA-256 |
|---|---|
model.safetensors | a0405c89c8a8a4dc78b04b1b3850e3a4b7c00f0a18dbab171d88c39693f59b4b |
Apache-2.0, the licence of the base model Qwen/Qwen3-1.7B-Base. The dataset is published under
CC BY 4.0, the code under MIT.
The code, the settings and the pipeline, at the version of these weights: https://github.com/BenoitJT-GIRARD/clinical-triage-llm/tree/v2.0.0. The adapters and the intermediate merged model are not published: the pipeline rebuilds them from the dataset, at the settings these weights were trained with.
| n/a |