lqh 0.4.11

Liquid Harness – customize Liquid AI foundation models into task-specific models.

lqh — Liquid Harness

From zero to a fine-tuned model in under an hour.

Liquid Harness is a terminal agent that turns a plain-English description of your task into a small, fast, task-specific Liquid Foundation Model. You describe the problem; the agent interviews you, writes the specification, generates and curates training data, fine-tunes, evaluates, and iterates until the model beats baseline. No ML experience required.

[!IMPORTANT] ⚠️ Closed beta — visit lqh.ai to request access.

pip install lqh
lqh

---

🤔 Why would I want this?

Large general-purpose models are expensive, slow, and overkill for most production tasks. A 350M–1.2B model fine-tuned on your task is cheaper, faster, runs on-device, and often more accurate — but getting there normally requires an ML team: data pipelines, judges, training loops, eval harnesses.

Liquid Harness collapses all of that into a conversation:

1. 💬 Describe your task — the agent asks clarifying questions and writes a spec.
2. ☕ Let it work — it generates synthetic data, scores every sample against a rubric, filters, runs baselines, fine-tunes, and evaluates.
3. 📦 Get a checkpoint — a model that measurably beats the baseline on your task, ready to deploy.

💡 What can you build?

A few things people fine-tune with lqh:

Use case	You tell the agent…
🎧 Support-reply rewriter	"Rewrite draft replies into our brand voice: warm, concise, never over-promising."
🧾 Structured extraction	"Turn free-form purchase emails into strict JSON with vendor, amount, and date."
🚦 Classifier / router	"Label incoming tickets as billing, bug, or feature request — with high recall on billing."
🖼️ Vision Q&A	"Here's a folder of product photos — I need a model that answers questions about defects."
📱 On-device assistant	"A tiny model that summarizes meeting notes offline on a phone."

Point it at a folder of raw, unlabelled images and it can build a vision fine-tune too — the data pipeline synthesizes the question-answer pairs for you.

🚀 Quickstart

pip install lqh
mkdir my-task && cd my-task
lqh

Inside the TUI:

> /login          # authenticate with lqh.ai (one-time)
> I want a model that rewrites support replies into our brand voice.

That's it. The agent takes it from there — it interviews you about requirements, writes SPEC.md, and offers to run the pipeline stage by stage. You can interject, inspect data samples, or change direction at any point.

[!TIP] Run /hf_login (or set HF_TOKEN) to enable private HuggingFace dataset access and publishing.

🔁 The pipeline

One command runs all nine stages. Each is a real artifact you can inspect, stop at, or hand off.

spec → rubric → data gen → filter → baseline → SFT → DPO → eval → checkpoint

$ lqh --auto ./my-task
[stage: rubric]            writing scorer from spec
[stage: data_gen_draft]    5 samples generated, all valid
[stage: filter_validation] 1,427 / 2,000 kept
[stage: sft_initial]       score 6.8/10  (baseline 4.1)
[stage: dpo]               iter 3/5, score 7.4/10
[final: success]           DPO checkpoint beats baseline by +3.3

✨ Features

💬 Fully agentic

You chat, the agent works. It captures requirements through dialogue, manages the specs, and drives every downstream stage end-to-end.

📝 Specify in plain English

No DSL, no boilerplate, no ML jargon. Just describe what you want — the agent turns it into structured specs you can refine over time.

🧪 Synthetic data, scored & filtered

The agent authors a per-task data generation pipeline, generates samples concurrently on LQH Cloud, and scores each one with an LLM judge against your rubric. The dataset that hits training is already curated.

🖼️ Vision fine-tuning

Bring a folder of raw images — no labels needed. The agent synthesizes an image-question-answer dataset and fine-tunes the LFM2.5-VL vision models on it.

🏋️ Train anywhere

Eval and data generation run on LQH Cloud. Training runs locally on your own GPUs, or hands off to a remote machine over SSH — including SLURM clusters — with dataset sync handled for you.

🤖 Hands-off --auto mode

Point lqh at a directory with a spec and walk away. It either delivers a checkpoint that beats baseline or returns an explicit failure with the reason — never a hang, never a prompt.

🤗 HuggingFace integration

Push and pull datasets from the Hub, publish checkpoints, and convert to GGUF for deployment.

🖥️ Interactive TUI

Guide the agent, visualize progress, and inspect dataset samples — all from one terminal session with a slash-command palette and a live status bar.

📦 Project-as-directory

Any directory is a project — fully git-compatible, so you can version, branch, and collaborate on specs, datasets, and runs like any other code. cd to switch projects.

🧬 Models

Fine-tunes the LFM2.5 family — pick a size for your task, from tiny on-device to MoE:

Size	Best for
230M / 350M	Very simple tasks, extreme on-device constraints
1.2B ⭐	Recommended starting point for most tasks
2.6B / 8B-A1B (MoE)	More complex tasks
24B-A2B	Only when the task clearly calls for it
LFM2.5-VL 450M / 1.6B	Vision (image + text) tasks

Base, instruct, and thinking variants are available; the agent recommends the right starting size and variant for your task and steps up if fine-tuning struggles.

⌨️ Slash commands

Command	What it does
/login	Log in to lqh.ai
/hf_login	Store a HuggingFace token for cloud jobs
/spec	Start specification capture
/datagen	Start data generation
/validate	Start data validation
/train	Start training (requires torch)
/eval	Start evaluation
/prompt	Start prompt optimization
/resume	Resume a previous conversation
/clear	Start a fresh conversation
/reconnect	Retry a failed network/API operation
/help · /quit	Show commands · exit

🤖 Auto mode

For CI, batch jobs, or when you just want a result:

lqh --auto ./my-task                # runs the full pipeline against ./my-task/SPEC.md
lqh --spec "use the smallest base model"   # sticky run-time directive (works in both modes)

Auto mode requires an existing SPEC.md (write one interactively first, or by hand). It runs rubric → data gen → filter → baseline → SFT → DPO → report without ever prompting, and always terminates with an explicit success or failure.

📁 Your project is just a directory

cd into any directory and run lqh — the agent reads what's there to understand current state. No init command, no project marker file.

my-task/
├── SPEC.md              # the heart of the project: what you want the model to do
├── other_specs/         # additional specs for edge cases or sub-requirements
├── data_gen/            # generated pipeline scripts
├── evals/               # eval definitions and results, versioned (v1/, v2/, ...)
├── datasets/            # generated and curated datasets as parquet (v1/, v2/, ...)
├── runs/                # training runs with checkpoints, logs, and configs
└── .lqh/                # conversation logs and permissions (add to .gitignore)

Everything is plain files — specs are markdown, pipelines are Python, datasets are parquet. Edit SPEC.md directly any time; the agent picks up your changes on the next turn.

🔧 Requirements

• Python 3.11+
• A Liquid Harness account (request access)
• Optional: torch + transformers for local fine-tuning
• Optional: HF_TOKEN for HuggingFace dataset sync

🗺️ Roadmap

Things we're actively building. Open an issue if you want to weigh in.

• Quantized evals — run local evaluation on the quantized artifact (llama.cpp) so we measure exactly what ships. GGUF export already works; the eval loop on top is in progress.
• Quantization-aware training (QAT) — train against quantization noise so the deployed quantized model matches the full-precision score.
• Sub-agent spawning — parallel sub-agents for independent subtasks (drafting multiple data pipelines, running evals concurrently).
• Hosted training API — train on LQH Cloud without bringing your own GPUs.

🤝 Contributing

Please read CONTRIBUTING.md before opening a pull request. Random PRs will be rejected — open an issue first and agree on the approach with a maintainer; security hot fixes are the one exception. See the contribution policy for details.

---

Made with care by Liquid AI.