⚡ Finetune Studio

FEATURES

Every stage of the local-LLM pipeline, wired together with a shared project concept. Pick what you need — most people start with one tab and grow.

AGENTIC DATA PREP TOOL-CALLING

A local model mines your documents for training data: list_sources → read_source → create_qa_pairs, with every tool call rendered inline so you can audit exactly what the AI did before it touches your dataset.

LoRA / QLoRA TRAINING

Point at a JSONL or your mined dataset, pick a base model, hit train. Live loss curves, VRAM-aware config, adapter merge, and a one-click bake to GGUF.

RAG CORPORAS

Ingest PDF / DOCX / CSV / code / images (OCR), chunk, embed, index. Search playground with top_k, threshold and reranker tuning. SSE progress the whole way.

FILE LIBRARY

SHA-256 dedup, MIME-auto-routed immutable raws, your own folders, versioning, and a 7-day soft-delete trash. Files are data; the DB is just the index.

BENCHMARKS

MMLU · HellaSwag · ARC · TruthfulQA · GSM8K · Winogrande + tool-calling evals. Side-by-side base-vs-trained tables, exportable JSON.

LOCAL INFERENCE

GGUF (llama.cpp) or safetensors (transformers), vision models included. Idle auto-unload frees VRAM; per-chat temperature, top-p and system prompt.

SELF-HEALING UPDATES

./update.sh or Settings → Apply update: pull, venv repair, dep sync, migrations, restart — with a live log stream in the UI.

VRAM DIAGNOSTICS

Out-of-VRAM? The load error names the culprit and its memory: “needs ~19.8 GB, only 10.3 free; top consumers: comfyui(pid 2063394)”.

SCREENSHOTS

Terminal-grade UI: tmux-style session bar, phosphor-green notches, animated sprites, command palette on Ctrl+K. Live on any modern browser at :7860.

Dashboard with host resources and project cards — **DASHBOARD** — host resources (RAM/VRAM bars), your projects, quick-create.

Data prep with file library and prepped Q&A — **DATA PREP** — file library, prep jobs, and the approved Q&A table mined by the agent.

Project chat with Test/Agent modes — **PROJECT CHAT** — Test vs Agent modes; amber banner when no model is online.

Inference page with loader settings — **INFERENCE** — pick a model, tune the loader (ctx, GPU layers, batch), chat.

Training page with live loss — **TRAINING** — LoRA config, live loss curve, the CpuChip sprite flickers as the GPU works.

Benchmarks comparison — **BENCHMARKS** — base vs trained, score bars fill live.

Settings with Server and Hosting card — **SETTINGS** — debug info, storage paths, Server & Hosting (port, CORS, trusted hosts, proxy), and the self-healing Updates card.

THE PIPELINE

From raw documents to a benchmarked, chat-ready model — one project, one flow:

upload ──► file library      sha256 dedup · MIME auto-folders · versioning · 7-day trash
   └─► parse + chunk      pdf · docx · csv · xml · images(OCR) ── ~500-token chunks
        └─► Q&A mining       batch prep job  OR  agentic tool-calling chat
             └─► review           approve / reject / edit pairs ──► JSONL datasets
                  └─► train           LoRA / QLoRA / full · live loss · VRAM profiles
                       └─► merge + bake     adapter → merged → GGUF (quantized)
                            └─► benchmark        MMLU · GSM8K · tool-calling · base-vs-trained
                                 └─► chat             single-model engine · vision · RAG context

WHY NOT JUST…?

None of the pieces are impossible by hand. The point is that all of them are now in one place, with your data in ~/.finetune-studio/ where you can find it.

What you'd do manually	What Finetune Studio does
Five terminal windows + a Notion page of "things to remember"	One browser tab, one project concept
Hand-craft chunking + embedding loops	Click Build — SSE progress the whole way
Write your own MMLU eval harness	Pick the suite, click Run, get a comparison table
Lose an afternoon to CUDA install issues	One ./install.sh — GPU-aware wheels, self-repair
“Failed to load model from file” (useless)	Actionable errors: needs 19.8 GB, 10.3 free, comfyui is holding it

QUICKSTART

Linux / macOS / Windows(WSL2). Auto-detects your GPU and picks the right CUDA wheels.

$ git clone https://github.com/GenorTG/finetune-studio.git
$ cd finetune-studio
$ ./install.sh                    # GPU detection, venv, wheels, llama.cpp CLI
$ source .venv/bin/activate
$ fts web --host 0.0.0.0 --port 7860
# open http://localhost:7860 — an 8-step tour walks you through it

⚡ FINETUNE STUDIO▮