Early access · Built for AI teams

Fine-tune models.
Ship better AI.

Adapt an open model to your support workflow. Explore LoRA training and policy-specific evaluation, with hosted training and API deployment on the product roadmap.

Open model foundations Your data. Your adapters.
Model pipelineWORKSPACE / 01
01 / PREPAREYour domain. Your dataset.

conversations.jsonl CSV · JSONL

Dataset validation & train / eval split
02 / ADAPTOpen model → your model

LoRA / QLoRA NVIDIA GPU

Evaluate checkpoints against a baseline
03 / DEPLOYReady for your application.

POST /v1/chat/completions

YOUR DATA → YOUR MODEL → YOUR APIDesigned for NVIDIA GPU accelerated training and inference
THE WORKFLOW / IN 30 SECONDS

From support data to your own model.

Watch an illustrated guide to preparing data, adapting an open model, and evaluating it before a planned API release.

Illustrated product workflow. Shows the intended experience. Hosted training and API deployment are planned.

See the actual local experiment
Read the video transcript
  1. Bring your support data. Prepare resolved support tickets in JSONL or CSV. Validate labels and keep a separate evaluation set.
  2. Adapt an open model. Choose a base model and train a LoRA adapter on your examples. The adapter captures task specific behaviour.
  3. Evaluate before release. Check structured outputs, ticket routing and security escalation. Only a candidate that passes your policy checks should move forward.
  4. Connect it to your application. The planned inference API returns a structured decision to your support application. Hosted API deployment remains on the roadmap.
Explore the interactive platform preview
YOUR NEXT MODEL STARTS HEREInteractive product preview
Workspace / Training jobs Preview environment
PROJECT / CUSTOMER SUPPORT

Monitor your training run.

Sample project
support-assistant-v1Meta Llama 3.1 8B · QLoRA
Training
Training loss0.284 ↓ 82.4%
Epoch2 / 3
GPUNVIDIA A100
Elapsed time00:24:18
Training loss train_loss eval_loss
0200400600800 steps
Training in progress68% complete
LoRA & QLoRA training Your data stays yours API deployment
REAL DEVELOPMENT. VISIBLE LIMITATIONS.

A completed LoRA experiment you can inspect.

64 synthetic training tickets. A saved adapter. All 16 test outputs. The candidate returned valid JSON but missed the routing threshold, so promotion was blocked.

16 / 16Valid JSON outputs
12 / 16Complete decisions correct
BlockedExploratory promotion gate

Local Apple MPS run. Synthetic data only. NVIDIA execution and customer validation remain unverified.

START WITH THE OPEN MODEL THAT FITS YOUR TASK.
M mistralai

Planned model ecosystem · Model availability and licenses vary

✳ PURPOSE BUILT FOR YOUR AI

Custom AI, without the infrastructure backlog.

Keep datasets, training runs, evaluations, and endpoints together. Give your team a repeatable path from a first experiment to an application integration.

EFFICIENT TRAINING

Train without the GPU setup.

Adapt open models with LoRA and QLoRA. Spend your time improving results, not configuring training infrastructure.

MODEL OWNERSHIP

Keep control of your models.

Build on open model foundations with your own datasets and adapters. Keep control of what makes your AI different.

SECURITY BY DESIGN

Trace every experiment.

Connect each checkpoint to its dataset, training configuration, and evaluation. Keep the information your team needs to reproduce a result.

API FIRST

Integrate through an API.

Move from a model checkpoint to an API endpoint, with an integration pattern your developers already know.

✳ A SHORTER PATH TO PRODUCTION

From dataset to deployed model.

Prepare examples, adapt a foundation model, and validate the result before you put it in front of users.

See how it works
01

Bring your data

Upload your dataset, validate the format, and create your training and evaluation splits.

support_conversations.jsonl
02

Make the model yours

Choose an open model, configure LoRA, and follow every training step in your workspace.

method: qlora · epochs: 3
03

Deploy. Connect. Build.

Evaluate your model and connect your application to a dedicated inference endpoint.

POST /v1/chat/completions
SERIOUS COMPUTE. SIMPLE EXPERIENCE.

GPU acceleration.
Developer control.

The platform architecture targets NVIDIA GPUs for CUDA accelerated training and inference. Configure model runs in a workspace instead of assembling GPU environments for every experiment.

CUDALoRA / QLoRAGPU inference
Explore the architecture

NVIDIA NIM, NeMo, and TensorRT-LLM integrations are on the roadmap.

GPUACCELERATED COMPUTE
TRAINING → EVALUATIONINFERENCE → YOUR APPLICATION
✳ BUILT FOR YOUR ACTUAL WORKLOAD

Teach a model what makes your product different.

Fine-tuning is useful when you need consistent behavior, a specific output format, or domain language that a generic prompt cannot reliably deliver.

01
EXTRACTION & CLASSIFICATION

Reliable structured outputs

Train on examples of your schema so applications can work with predictable classifications, fields, and response formats.

Explore the workflow
02
CUSTOMER EXPERIENCE

Support that knows your product

Adapt an open model to the tone, terminology, and resolution patterns in your approved support conversations.

Explore the workflow
03
DOMAIN ADAPTATION

Models for a specialist domain

Use curated examples to improve task behavior in your domain. Combine with retrieval when your application needs current facts.

Explore the workflow
OPEN FOUNDATIONS. YOUR OWN DIRECTION.

Why we’re building Forgedata
A LITTLE MORE CLARITY

Good questions.
Straight answers.

Have something else in mind?

Talk to the team
What is Forgedata?

Forgedata is an AI product platform being built to bring dataset management, parameter efficient fine-tuning, model evaluation, and inference deployment into one workspace. Request early access to discuss current availability.

Which models can I fine-tune?

The planned model catalog includes the Llama, Mistral, Qwen, and Gemma families. Specific model versions and available GPU configurations will be confirmed during early access. Base model licenses continue to apply.

What dataset formats are supported?

The platform workflow is designed for JSONL and CSV datasets, with schema validation and training and evaluation splits before a job starts.

Do I own my fine-tuned models?

Model ownership and export are core product goals. Your rights to adapters and weights remain subject to the base model license and the final platform agreement.

How does Forgedata use NVIDIA GPUs?

The architecture targets CUDA accelerated training on NVIDIA GPUs. TensorRT-LLM optimization, NVIDIA NIM deployment, and NeMo evaluation are planned integrations, on the platform roadmap.

Can I use the inference API today?

Request early access to confirm endpoint availability for your workload. The interactive preview shows the intended API workflow; it does not provision a live endpoint.

YOUR NEXT CHAPTER IN AI

From your first dataset.
To your next release.

Bring a use case. Explore a model workflow built for your team.

Request access