Spaces:
Running on Zero
A newer version of the Gradio SDK is available: 6.28.0
title: MLOL — MultiDomain LLM Optimisation Lab
emoji: 🧪
colorFrom: green
colorTo: gray
sdk: gradio
sdk_version: 5.34.2
app_file: app.py
pinned: false
license: apache-2.0
🧪 MLOL — MultiDomain LLM Optimisation Lab
A complete LLM fine-tuning and optimisation laboratory on a Hugging Face Space:
Choose Base Model → Upload Dataset → Validation & Cleaning → Configure →
Hardware Recommendation → Baseline Eval → Fine-Tune → Post-Eval →
Comparison → Optimisation Analysis → Performance Certificate → Report → Deploy
Design: the Space is the control plane. Training runs on the right backend per the routing engine — ZeroGPU for bounded demos (≤1.5B), pinned Colab export packages (free), or HF Jobs (managed). All experiment state persists to a private Hub dataset repo; the Space is stateless and restart-safe. Full blueprint: docs/MASTER_SPEC.md · usage: docs/USER_GUIDE.md.
Modules
Home · Tier 1 — General Fine-Tuning Lab · Tier 2 — Domain Foundry (premium; 11 domains) · Evaluation Lab (seeded samples, CIs, paired significance tests) · Reports (9-section Model Performance Certificate, PDF/CSV/JSON) · Adapter Library · Hardware Advisor · Documentation · AI Research Assistant (bottom-right; General/Experiment/Hardware/Report modes, config-driven providers).
Model and provider catalogues are pure configuration (configs/*.yaml) —
add newly released models with zero code changes.
Space setup
- Hardware: ZeroGPU (PRO) recommended; works on CPU with reduced function.
- Secrets:
HF_TOKEN(gated models + Hub persistence),PREMIUM_ACCESS_CODES(Tier 2 gate; unset = open dev mode). - Premium / custom domains: finpy07@gmail.com.
Repo layout
app.py # UI (control plane only)
configs/ # models, providers, domains, hardware, limits
src/schemas.py # Pydantic configs + experiment manifest/state machine
src/config_loader.py # startup schema validation
src/services/ # persistence, dataset_prep, routing, training
# backends (mock/zerogpu/colab/jobs), evaluation,
# reporting, assistant
src/inference/engine.py # base+adapter inference
src/data/, src/training/ # legacy FinLLM pipeline (reused as training payload)
docs/MASTER_SPEC.md # the binding blueprint (v1.1)
Research platform — outputs are not investment, legal, medical, or other professional advice.