Model Card for Alz-SumAgent
Alz-SumAgent is a fine-tuned sequence-to-sequence summarization model, built on facebook/bart-large-cnn, that condenses multi-document Alzheimer's disease research evidence — retrieved PubMed abstracts and key findings — into concise, evidence-grounded summaries. It powers the "Deep" analysis mode of AlzDetect AI, a two-call agentic pipeline that pairs this summarizer with Claude for richer, multi-step research synthesis.
Model Details
Model Description
Alz-SumAgent was fine-tuned from facebook/bart-large-cnn (itself pre-trained on CNN/DailyMail news summarization) on Alzheimer's disease research abstracts and summaries, adapting it from general news summarization to dense biomedical language (e.g. amyloid-beta pathology, tau phosphorylation, APOE4 genetics, biomarker terminology).
In production, it is called as the first stage of a two-call agent (generation/summarizer_agent.py): retrieved PubMed chunks are first condensed by Alz-SumAgent, then the condensed summary is passed to Claude for final citation-grounded synthesis. This two-step design allows the pipeline to reason over a larger evidence base per query than a single-call approach could fit in context.
- Developed by: Tamil Priya Balaji (tpriyadata)
- Model type: Sequence-to-sequence transformer (BART architecture), fine-tuned for summarization
- Language(s): English
- License: MIT
- Finetuned from model:
facebook/bart-large-cnn
Model Sources
- Repository: tpriyadata/alz-sumagent
- Demo: oviya08/AlzSumAgent — select "Deep" analysis mode to invoke this model
Uses
Direct Use
Summarizing biomedical/clinical text related to Alzheimer's disease and related neurodegenerative research — e.g. condensing multiple paper abstracts or retrieved evidence chunks into a shorter synthesis. Best suited to text in a similar domain and register to its fine-tuning data (peer-reviewed research abstracts).
Downstream Use
Used as the first stage of a retrieval-augmented generation (RAG) pipeline: retrieved evidence chunks → Alz-SumAgent condenses → Claude synthesizes a final, citation-grounded answer. See AlzDetect AI for the live implementation.
Out-of-Scope Use
- Not for clinical decision-making. This model summarizes research literature; it does not provide diagnoses, treatment recommendations, or patient-specific guidance.
- Not trained or evaluated on patient-level clinical data. Training data is limited to publicly available PubMed abstracts. This model has never been trained on, fine-tuned with, or exposed to ADNI (Alzheimer's Disease Neuroimaging Initiative) or any other Data Use Agreement–restricted dataset. It is safe for use in DUA-compliant pipelines specifically because it only ever processes public literature, never participant-level clinical records.
- Not intended for general-purpose summarization outside biomedical/research text — performance on other domains is untested.
Bias, Risks, and Limitations
- Inherits any topical or publication bias present in the PubMed literature it was fine-tuned on (e.g. more heavily studied mechanisms like amyloid pathology are likely over-represented relative to less-funded research areas).
- As with any abstractive summarization model, it can occasionally compress or omit nuance from source text. In the production pipeline, this risk is mitigated by Claude's downstream citation-grounding step, which is instructed to answer only from retrieved source text.
- Not evaluated for factual consistency (faithfulness) in isolation — faithfulness metrics reported for the AlzDetect AI pipeline reflect the full two-stage system, not this model alone.
Recommendations
Use as an intermediate summarization stage within a citation-grounded RAG pipeline rather than as a standalone source of factual claims. Downstream consumers of its output should still verify claims against original cited sources.
How to Get Started with the Model
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
tokenizer = AutoTokenizer.from_pretrained("oviya08/Alz-SumAgent")
model = AutoModelForSeq2SeqLM.from_pretrained("oviya08/Alz-SumAgent")
text = "Your Alzheimer's research abstract or evidence chunk here..."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=1024)
summary_ids = model.generate(**inputs, max_length=150, min_length=40, length_penalty=2.0, num_beams=4)
print(tokenizer.decode(summary_ids[0], skip_special_tokens=True))
Training Details
Training Data
Fine-tuned on Alzheimer's disease research abstracts and summaries sourced from PubMed (public, peer-reviewed biomedical literature only). Part of the same 19,637-paper PubMed corpus used by the AlzDetect AI retrieval pipeline. No ADNI or other DUA-restricted data was used in training.
Training Procedure
Preprocessing
[More Information Needed]
Training Hyperparameters
- Training regime: [More Information Needed]
Speeds, Sizes, Times
- Model size: 406M parameters (verified via config inspection: hidden_size=1024, 12 encoder layers, 12 decoder layers)
- Tensor type: F32 (Safetensors)
Evaluation
Testing Data, Factors & Metrics
Evaluated as part of the full Alz SumAgent RAG pipeline (retrieval → summarization → Claude synthesis) using RAGAS-style metrics (hit rate, faithfulness, keyword coverage) rather than in isolation. See the AlzDetect AI Space for current pipeline-level evaluation results.
Results
[More Information Needed — standalone summarization metrics, e.g. ROUGE, to be added]
Environmental Impact
- Hardware Type: [More Information Needed]
- Hours used: [More Information Needed]
- Cloud Provider: [More Information Needed]
- Compute Region: [More Information Needed]
- Carbon Emitted: [More Information Needed]
Technical Specifications
Model Architecture and Objective
BART (Bidirectional and Auto-Regressive Transformer) encoder-decoder architecture (bart-large-cnn config: 1024 hidden size, 12 encoder + 12 decoder layers), fine-tuned for the sequence-to-sequence summarization objective on biomedical abstracts, starting from CNN/DailyMail news-summarization pretrained weights.
Compute Infrastructure
[More Information Needed]
Citation
If you use this model, please cite the Alz SumAgent project:
BibTeX:
@misc{balaji2026alzsumagent,
author = {Balaji, Tamil Priya},
title = {Alz-SumAgent: A Fine-Tuned BART Summarizer for Alzheimer's Disease Research},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/oviya08/Alz-SumAgent}}
}
Model Card Authors
Tamil Priya Balaji