This model is a low-rank adapter (LoRA) fine-tuned on top of togethercomputer/Llama-4-Scout-17B-16E-Instruct_bnb_4bit using the Adaption Labs orchestration platform. It is engineered specifically for the AutoScientist Challenge (Healthcare Track) to achieve state-of-the-art visual-textual reasoning across complex multi-modal medical diagnostics.

Model Details

Model Description

Developed by: Asad Ullah Dogar
Model type: Mixture-of-Experts (MoE) Vision-Language Model Adapter (PEFT/LoRA)
Language(s) (NLP): English (100% Analytical Clinical Register)
License: Apache-2.0
Finetuned from model: togethercomputer/Llama-4-Scout-17B-16E-Instruct_bnb_4bit

Model Sources

Repository: https://huggingface.co/asadullahdogarr/adaption-llama-4-scout-17b-mmmed-autoscientist-gold
Platform Orchestration: Adaption Labs Engine

Uses

Direct Use

This model is fine-tuned to solve high-entropy medical visual question answering (VQA) tasks. It accepts complex multi-modal case files containing visual diagnostic artifacts (such as X-rays, CT scans, ultrasounds, and histopathology profiles) accompanied by dense text narratives, and generates structured, multi-step clinical reasoning chains concluding with a definitive choice path execution.

Out-of-Scope Use

This model is an experimental checkpoint built for a benchmark hackathon challenge. It is not certified for diagnostic use in real-world clinical settings, primary triage, or autonomous patient treatment planning without certified human-in-the-loop medical supervision.

Bias, Risks, and Limitations

The model reflects patterns present in its clinical fine-tuning matrix. It exhibits strong capability in structured clinical text alignment but may inherit standard limitations associated with quantized low-rank multi-modal adapters, such as sensitivity to alternative formatting layouts if decoupled from its specific prompt configuration template.

Recommendations

Users evaluating this checkpoint should ensure input text prompts follow the structured pre-concatenation format: "Clinical Case Analysis:\n{Question}\nOptions:..." to maximize multi-agent reasoning chain activation.

How to Get Started with the Model

You can load this model using the standard Hugging Face Transformers and PEFT pipelines:

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel, PeftConfig

base_model_id = "togethercomputer/Llama-4-Scout-17B-16E-Instruct_bnb_4bit"
adapter_model_id = "asadullahdogarr/adaption-llama-4-scout-17b-mmmed-autoscientist-gold"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id, 
    torch_dtype=torch.float16, 
    device_map="auto"
)

model = PeftModel.from_pretrained(base_model, adapter_model_id)