Gretel AI Shut Down. Here's What That Means for Training Data — And What Comes Next.

In February 2026, Gretel AI — one of the most prominent companies in synthetic data generation — was archived. Their GitHub organization was marked as no longer maintained. The platform that thousands of teams relied on for generating synthetic datasets, including training data for LLM fine-tuning, effectively ceased operations.

If you were a Gretel user, you're now looking for alternatives. If you weren't, the shutdown still matters — because it reveals something important about where the training data market is heading and where it's failing.

What Gretel Did

Gretel built tools for generating synthetic data. Their platform covered a wide range of use cases: creating privacy-safe versions of tabular datasets, generating synthetic text for model training, and producing QA pairs with automated scoring for conformance, quality, toxicity, bias, and accuracy.

They had integrations with major cloud platforms including Azure AI Foundry. They offered both open-source libraries and commercial services. For teams building LLMs or fine-tuning models, Gretel was one of the few platforms that could generate structured training data at scale.

Their Navigator product was specifically designed for creating custom question-answer pairs — the exact format needed for supervised fine-tuning. For many teams, Gretel was the closest thing to an automated training data pipeline that existed.

And now it's gone.

Why This Matters

The timing is significant. We're in a moment where enterprise fine-tuning demand is accelerating rapidly. Organizations across insurance, legal, healthcare, and financial services are realizing that generic foundation models don't understand their domain-specific documents, terminology, and reasoning patterns. Fine-tuning on proprietary data is the path to models that actually work for specialized tasks.

But fine-tuning requires training data. Specifically, it requires high-quality instruction-response pairs that teach the model how to reason about your domain. Creating this data has always been the bottleneck — and Gretel was one of the few companies offering an automated solution.

With Gretel gone, the market has a gap. And the remaining options all have significant limitations.

The Current Landscape

If you need training data for LLM fine-tuning in March 2026, here's what's available:

None of these options provide what Gretel offered at its best: an accessible, structured platform for generating quality training data. And none of them are specifically designed for the use case that's growing fastest — turning proprietary enterprise documents into domain-specific fine-tuning datasets.

What the Market Actually Needs

The teams we talk to — in insurance, legal, compliance, and enterprise operations — describe the same problem consistently. They have large collections of domain-specific documents. They know fine-tuning would make their AI dramatically better. But the process of turning those documents into training data is either too expensive, too slow, too technically complex, or requires them to compromise on data privacy.

What they need is a system that can ingest their actual documents — not generic web text, but insurance policies, legal contracts, regulatory filings, internal manuals, codebases — and produce structured training pairs that capture the domain's reasoning patterns.

Not just simple factual extraction ("What does Section 3 cover?") but conditional reasoning ("If the policyholder's revenue exceeds $50,000, does this coverage still apply?"), exclusion logic ("What is explicitly NOT covered under this provision?"), and cross-document synthesis ("How do these two policies interact when both contain overlapping liability clauses?").

The pairs need quality control — not every AI-generated question is worth training on. Ephemeral data should be filtered out because it doesn't teach domain knowledge. Boilerplate and duplicate content should be detected and excluded before any pairs are generated. And a human expert should be able to review every pair and decide whether it meets the bar.

Finally, the entire process should run locally. If you're working with confidential insurance policies or legal contracts, the training data pipeline should have the same privacy guarantees as the documents themselves.

Where FORGE Fits

This is the specific problem we built FORGE to solve.

FORGE is an add-on to Librarian, our local AI document assistant. Librarian handles the document ingestion — parsing PDFs, Word documents, Excel files, and code across 20+ languages, then chunking and indexing them in a local vector database with semantic search.

FORGE reads those indexed chunks and generates training pairs automatically. But not randomly. It uses a structured approach:

The output is a clean JSON dataset in Alpaca format (instruction, input, response), ready to drop into any fine-tuning pipeline — Unsloth, Hugging Face, or any SFT toolchain.

And everything — ingestion, generation, review, and export — runs locally on your hardware. Documents never leave your machine.

Beyond Training Data

One thing Gretel never offered was the next step: actually fine-tuning the model.

FORGE doesn't stop at dataset generation. Once you have a reviewed, quality-gated dataset, FORGE can pull a base model from HuggingFace, run LoRA/QLoRA fine-tuning on your local GPU, and deploy the resulting model directly into Librarian — all within the same system.

The full pipeline looks like this:

No separate tools. No cloud services. No Python scripts to stitch together. One system, from raw documents to a deployed domain-specific model.

Who This Is For

FORGE isn't for everyone. It requires an NVIDIA GPU with at least 8 GB of VRAM for training (24 GB+ recommended for serious work with larger models). It's designed for professionals and teams who work with domain-specific documents and need AI that genuinely understands their field — not a general-purpose chatbot that gives surface-level answers.

If you're an InsurTech company building AI tools for the insurance industry, a law firm that needs AI trained on your contract library, a compliance team that wants AI fluent in your regulatory framework, or an enterprise with proprietary documentation that a generic model can't handle — FORGE is built for your use case.

If you were using Gretel for domain-specific training data and need an alternative that runs locally and goes beyond data generation into actual fine-tuning, we'd welcome the chance to show you what FORGE produces.

Getting Started

FORGE is available as an add-on to any Librarian subscription. You can start with a free 30-day trial of Librarian, index your documents, and see the quality of the search and retrieval before adding FORGE for training data generation and fine-tuning.

If you want to see what FORGE produces before committing, send us a few sample documents. We'll generate a demo dataset and send you the JSON output — no commitment, no cost. You can evaluate the pair quality yourself before deciding.