srikarjy025/lipidos-phi3-domain-adapt-merged

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:4kPublished:Aug 16, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

srikarjy025/lipidos-phi3-domain-adapt-merged is a 4 billion parameter Phi-3.5 Mini model, domain-adapted by srikarjy025 for lipid biochemistry and Raman/IR spectroscopy literature. This merged model, trained with Unsloth, achieves a 19.5% lower perplexity on relevant scientific abstracts compared to its base model. It is specifically designed for use within a citation-grounded retrieval augmented generation (RAG) system to ensure factual accuracy and prevent hallucination in specialized scientific domains.

Loading preview...

LipidOS - Domain-Adapted Phi-3.5 Mini

This model, srikarjy025/lipidos-phi3-domain-adapt-merged, is a standalone merged version of microsoft/Phi-3.5-mini-instruct, specifically adapted for the scientific domains of lipid biochemistry and Raman/IR spectroscopy. Developed by srikarjy025 as part of the LipidOS project, it focuses on improving language understanding within these specialized fields.

Key Capabilities & Training

  • Domain Adaptation: Fine-tuned on 107,665 PubMed abstracts related to lipid biochemistry and Raman/IR spectroscopy, with 64,000 abstracts used for training.
  • Perplexity Improvement: Achieves a measured perplexity of 3.955 on a held-out dataset, representing a 19.5% reduction compared to the base Phi-3.5-mini-instruct model (4.911).
  • Citation Grounding: Designed to maintain citation-grounding capabilities, verified to avoid hallucinated evidence citations when used within its intended pipeline.
  • Merged Model: This is a merged model (base + QLoRA adapter), eliminating the need for separate peft loading.

Intended Use Case

This model is specifically engineered for use within a grounded-retrieval + citation-checking pipeline. It is not recommended for open-ended generation without attached retrieved evidence. Its primary strength lies in enhancing the accuracy and domain relevance of responses within RAG systems for scientific literature, particularly in the fields of lipidomics and vibrational spectroscopy.