mednli_laya

mednli_laya

Checks a medical claim against a piece of evidence and answers supported, unsupported or contradicted. A fine-tuned Laya model for English research on clinical-note text, not for clinical decisions.

Made by pavanmantha on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: mednli_laya is a Laya decision model tuned to check medical claims three ways. You give it a claim and a piece of evidence, and it answers supported, unsupported or contradicted. Pavan Kumar Mantha trained it from a multilingual Laya checkpoint on the MedNLI dataset.

Who it suits: It suits researchers and developers who work with English clinical-note text and want a small model that judges whether evidence backs a claim.

Title: Who might use it
Industry | Who uses it | How they use it | Example
Medical NLP research | Researchers | Test claim and evidence checking on clinical sentences | Comparing a note line with a short claim about a lab value
Health-tech prototyping | Developers | Prototype a claim checker for research use only | Flagging a claim that a note contradicts
Model evaluation | ML engineers | Study accuracy, calibration and resilience of a small classifier | Reading the benchmark tables and seed results

People who need clinical decisions, non-English text or long guideline passages can skip it.

Where to use it:

• Python: load the model with the laya library
• Windows, macOS or Linux: download the files with Python
• Research notebooks and scripts: score claim and evidence pairs one at a time

Title: Jobs it does
Use case | Who | What they get back
Claim checking | Researchers | One of three verdicts for a claim and its evidence
Benchmarking | ML engineers | A model to compare against the original Laya checkpoint
Prototype checks | Developers | A verdict they can inspect before building further

How to set it up: You need these first:

• Python on Windows, macOS or Linux
• The laya library, which the example code imports
• About 644 MB of storage for the F16 weights

1. Install the Hugging Face hub package.
2. Download the model files.
3. Install the laya library, then run the usage code.

pip install -U huggingface_hub
hf download pavanmantha/mednli_laya

How to use it: Load the model, define a question with the three answer criteria, pass the claim and the evidence, and read the verdict. Pass the claim before the evidence, as the model saw them during training.

import laya

agent = laya.load("pavanmantha/mednli_laya")
questions = {
  "verdict": {
    "type": "choice",
    "instructions": "Decide whether the cited text supports the claim.",
    "criteria": {
      "supported": "the cited text confirms the claim",
      "unsupported": "the cited text does not say enough to confirm or deny the claim",
      "contradicted": "the cited text states the opposite of the claim"
    }
  }
}
state = {"claim": "Patient has elevated Cr", "evidence": "Labs were notable for Cr 1.7 (baseline 0.5)."}
print(agent.predict(state, questions)["answers"]["verdict"])

Title: Calls used above
Tool | What it does
laya.load | Loads the fine-tuned model by its Hugging Face name
agent.predict | Takes the claim and evidence plus the question, and returns the answers

Example: You pass the claim that a patient has a raised creatinine level and a lab line showing a value well above baseline. The model returns one of its three labels: supported, unsupported or contradicted.

For newcomers and experts: If you are new to this, think of it as a referee that reads a claim and a note and says whether the note confirms it, says too little or says the opposite. Run the usage code once and change the claim to see the verdict move.

If you work in the field, it was tuned on MedNLI, where entailment, neutral and contradiction stand in for the three verdicts. On the 1422-item test split, accuracy rose from 0.4494 for the original to 0.7904, and the card reports it fixed 581 errors while adding 96. Claim-only accuracy of 0.4114 is above chance, which the card calls a known property of MedNLI. The option-order flip rate fell from 0.0816 to 0.0155.

Write a claim and an evidence sentence
Load the model with laya
Ask the verdict question with three criteria
Read supported, unsupported or contradicted

Good to know: The card reports results on the MedNLI test split only, where chance accuracy is 0.3333.

Title: Accuracy on the MedNLI test split
Chance: 0.3333
Original Laya: 0.4494
This model: 0.7904

• License: apache-2.0
• Cost to run: none named
• Data it reaches: the claim and evidence text you pass in, while training used the public MedNLI dataset
• Not for: clinical decisions, long guideline passages or non-English text, and it is not a medical device

On Hugging Face, 13 November 2026

• Task: text classification
• Library: laya
• Parameters: 322M
• Weights: about 644 MB in F16; about 161 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 16
• Likes: 8
• License: apache-2.0
• Languages: en
• Based on: convaiinnovations/laya
• Tags: laya, safetensors, claim-verification, medical, text-classification, model-index

Links

• Model: https://huggingface.co/pavanmantha/mednli_laya
• Files: https://huggingface.co/pavanmantha/mednli_laya/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download pavanmantha/mednli_laya

Then follow the usage in its model card. Nothing in this kit runs by itself.