Model_Fit_Finder / README.md
AdamTT's picture
Update README.md
9d9b8ca verified
---
title: Model Fit Finder
emoji: 👀
colorFrom: red
colorTo: red
sdk: gradio
sdk_version: 6.4.0
app_file: app.py
pinned: false
license: apache-2.0
short_description: Space that helps you choose the right type of NLP model
---
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
# Model Fit Finder (CPU)
**Model Fit Finder** is a decision-support Space that helps you choose the **right type of NLP model** and **concrete Hugging Face models** for your task — without training, without GPU, and without guesswork.
The Space is designed to reflect real-world AI engineering decisions rather than showcase a single model demo.
---
## What this Space does
The Space guides the user through a small set of practical questions and then:
* identifies the **appropriate model category** (instruction, QA, embeddings),
* ranks and recommends **at least 3 concrete Hugging Face models**,
* explains **why these models were selected**,
* adapts recommendations based on **language, compute budget, and priority**,
* optionally pulls **up-to-date models directly from Hugging Face Hub**.
All recommendations are **CPU-friendly** and suitable for lightweight prototyping and production planning.
---
## Supported NLP tasks
The Space currently supports three common NLP problem types:
### 1. Chat / instruction-following / generation
For tasks such as:
* chatbots
* summarization
* explanation
* instruction-based text processing
Recommended models are **instruction-tuned text-to-text or generative models**.
---
### 2. Question Answering from documents (extractive QA)
For tasks where:
* you have a document or text,
* answers must come strictly from that text,
* hallucinations should be minimized.
Recommended models are **extractive QA models** fine-tuned on datasets like SQuAD.
---
### 3. Semantic similarity / search / deduplication
For tasks such as:
* finding semantically similar texts,
* detecting near-duplicates,
* semantic search,
* retrieval for RAG pipelines.
Recommended models are **embedding (sentence similarity) models**.
---
## How recommendations are generated
Recommendations are **not static**. The Space uses a simple but explicit decision logic based on:
* **Data language**
* EN
* PL
* Mixed / multilingual
* **Compute budget**
* Low (fast, small models)
* Medium (allows larger, higher-quality models)
* **Priority**
* Speed
* Quality
* **Model source**
* Curated (hand-picked, stable baseline)
* HF Live (fresh models from Hugging Face Hub)
* Hybrid (curated + live)
Each candidate model is scored using heuristics such as:
* model size (small vs base),
* language coverage (English vs multilingual),
* suitability for the selected budget and priority,
* stability (curated vs live).
The Space always returns **a minimum of three models**.
---
## Hugging Face Live mode
When **HF Live** or **Hybrid** mode is enabled, the Space:
* queries the Hugging Face Hub using task-specific pipeline tags,
* ranks models by popularity (downloads),
* applies language and budget heuristics,
* caches results locally (with TTL),
* allows manual refresh via a **“Refresh HF cache”** button.
This prevents the Space from becoming outdated while keeping results stable and interpretable.
---
## What this Space is (and is not)
**This Space is:**
* a model selection assistant,
* a practical decision tool,
* CPU-only and cost-free,
* suitable for engineers, analysts, and ML practitioners.
**This Space is not:**
* a chatbot demo,
* a benchmark leaderboard,
* an automatic “best model” oracle.
Its goal is to help you make **better-informed model choices**, not to hide trade-offs.
---
## Example use cases
* *“Which embedding model should I use to detect semantically similar Revit Key Notes?”*
* *“I have a policy document and want reliable question answering without hallucinations.”*
* *“I need a lightweight instruction-following model for short summaries on CPU.”*
* *“Which models make sense for Polish or mixed-language text?”*
---
## Technical notes
* No model training is performed.
* No GPU is required.
* All logic runs on CPU.
* Model recommendations are based on metadata, heuristics, and Hugging Face Hub signals.
---
## Why this Space exists
Choosing the right model is often harder than using one.
This Space focuses on **model selection reasoning** — the part that usually lives only in engineers’ heads — and makes it explicit, inspectable, and reusable.