Jaymerry/mistral-7b-itis-taxonomy-slm-v3

<img src="https://cdn-uploads.huggingface.co/production/uploads/67c8730f98b8e1b9850bb24e/S6v6QzWqr5dvY-KS6V9Mf.png" width="400">

ITIS Taxonomy SLM (Mistral 7B Fine-Tuned)

A lightweight Small Language Model (SLM) fine-tuned for taxonomy classification, based on Mistral-7B and optimized for local inference (GGUF). 👉 Use this model as:

• —an assistant
• —a formatter
• —a lightweight offline tool

👉 Do NOT use it as:

• —a replacement for ITIS
• —a source of scientific truth without verification

<img src="https://cdn-uploads.huggingface.co/production/uploads/67c8730f98b8e1b9850bb24e/S_yqEGRhpoWKTVwI-bwsZ.png" width="800">

⚠️ Disclaimer

This model is not an authoritative source for full taxonomic lineage.

While it often produces well-structured outputs, individual lineage relationships may be incorrect.

Overview

This model is a domain-specific Small Language Model (SML) fine-tuned on structured data derived from the Integrated Taxonomic Information System (ITIS).

It explores how a relatively small model (~7B parameters) can perform on:

• —structured tasks
• —stable datasets
• —domain-specific reasoning

The goal is not to replace large models, but to demonstrate a lightweight, offline-capable alternative for constrained use cases.

Key Characteristics

• —Base model: Mistral 7B Instruct
• —Fine-tuning: QLoRA (4-bit)
• —Dataset: ~35K ITIS-derived samples
• —Quantization: GGUF (Q5_K_M)
• —Domain: Biological taxonomy
• —Designed for: offline / controlled environments

What This Model Does Well

This model performs best on short, structured outputs:

• —✅ Taxonomic rank identification
• —✅ Parent taxon identification
• —✅ Scientific ↔ common name mapping
• —✅ Basic validity checks
• —✅ Unknown taxon detection

👉 These tasks are highly reliable in practice

What Requires Caution

⚠️ Full Lineage Generation

Even when:

• —JSON is valid
• —structure looks correct

➡️ the taxonomy itself may be partially incorrect

Example risk:

• —correct genus
• —incorrect family

Benchmark (Strict Evaluation)

Key takeaway:

👉 Short tasks are reliable 👉 Long structured outputs require verification

Recommended Usage

High-confidence use cases

• —rank queries
• —parent queries
• —structured QA on known taxa
• —offline taxonomy assistants

How to Use

Option 1 — LM Studio

Download the .gguf file Load into LM Studio Use with prompt example below

Option 2 — llama.cpp

./main -m mistral-7b-itis-taxonomy-sml-v3.gguf -p "What is the lineage of Panthera leo?"

Option 3 — Python (recommended for tracking)

from huggingface_hub import hf_hub_download

model_path = hf_hub_download( repo_id="Jaymerry/mistral-7b-itis-taxonomy-slm-v3", filename="mistral-7b-itis-taxonomy-sml-v3.gguf" )

Example Prompt

➤ Taxonomic rank Input:

• —What taxonomic rank is Panthera leo?
• —Output:
• —Species

<img src="https://cdn-uploads.huggingface.co/production/uploads/67c8730f98b8e1b9850bb24e/RVCApHUHWASZS3tCCIn5N.png" width="400">

Small Language Model (SML) Perspective

This project demonstrates that:

Small models can be effective when trained on structured, stable data.

Benefits:

• —low compute cost
• —*offline usage*
• —reproducibility
• —controlled knowledge scope

This is particularly relevant for:

• —enterprise use cases
• —scientific tooling
• —embedded / local AI

Dataset

Source: Integrated Taxonomic Information System (ITIS) https://www.itis.gov/

Properties:

• —public domain
• —structured
• —relatively stable over time

Legal Notice

• —ITIS data is public domain
• —this model is a statistical transformation, not a database
• —outputs may be incorrect or incomplete
• —verification is required for critical use

Limitations

• —domain-restricted
• —English-only
• —not general-purpose
• —lineage may contain factual errors
• —imperfect refusal on validity

Notes on Downloads

For accurate download tracking, prefer using: hf_hub_download(...) Downloads via direct links or external tools may not be counted.

Author

Jeremy Banchet (Jaymerry)