NEWSovereign fine-tuning on isolated compute — 50 free GPU minutes

Own your weights.Deploy anywhere.

Fine-tune open-source models for your production data. In-isolation runs, zero-data retention, training-to-endpoint in seconds.

Start training free

MULTI-GPU CLUSTER•PYTORCH 2.5 + CUDA 12.4•100% EXPORTABLE WEIGHTS•ZERO DATA RETENTION

● TrainingMistral-7B-Instruct-v0.3•Run #4092

A100-80GB SXM4•16.4 GB / 80 GB•CUDA 12.4

Current Loss

0.434(eval: 0.412)

Step: 1,240 / 2,000

LR: 2.0e-5 (Cosine)

Throughput: 4,820 tok/s

AdamW (β₁=0.9, β₂=0.999, eps=1e-8)Convergence: -76.4% delta

Stream Telemetry

[00:01:23]Epoch 1/3 — Step 240/1200 — Loss: 1.842 — Perplexity: 14.2

[00:04:15]Epoch 1/3 — Step 480/1200 — Loss: 1.104 — Perplexity: 8.9

[00:07:02]Eval check: perplexity 8.12 → 5.43 (val loss: 0.982)

[00:11:45]Checkpoint saved: checkpoint-800 [adapter.safetensors 48MB]

[00:14:20]Epoch 2/3 — Step 1000/1200 — Loss: 0.512 — Perplexity: 3.8

[00:18:02]Final eval: perplexity 3.21 — Target convergence: 0.434

[00:18:15]Exporting LoRA weights to isolated artifact store... Done

[00:18:22]Zero retention: ephemeral GPU container flushed

[00:18:30]Serving endpoint online: https://api.langtrain.ai/v1

Listening for cluster events...

Model: mistral-7b-custom-v1

NEWSovereign fine-tuning on isolated compute — 50 free GPU minutes

Own your weights.Deploy anywhere.

Fine-tune open-source models for your production data. In-isolation runs, zero-data retention, training-to-endpoint in seconds.

Start training free

MULTI-GPU CLUSTER•PYTORCH 2.5 + CUDA 12.4•100% EXPORTABLE WEIGHTS•ZERO DATA RETENTION

● TrainingMistral-7B-Instruct-v0.3•Run #4092

A100-80GB SXM4•16.4 GB / 80 GB•CUDA 12.4

Current Loss

0.434(eval: 0.412)

Step: 1,240 / 2,000

LR: 2.0e-5 (Cosine)

Throughput: 4,820 tok/s

AdamW (β₁=0.9, β₂=0.999, eps=1e-8)Convergence: -76.4% delta

Stream Telemetry

[00:01:23]Epoch 1/3 — Step 240/1200 — Loss: 1.842 — Perplexity: 14.2

[00:04:15]Epoch 1/3 — Step 480/1200 — Loss: 1.104 — Perplexity: 8.9

[00:07:02]Eval check: perplexity 8.12 → 5.43 (val loss: 0.982)

[00:11:45]Checkpoint saved: checkpoint-800 [adapter.safetensors 48MB]

[00:14:20]Epoch 2/3 — Step 1000/1200 — Loss: 0.512 — Perplexity: 3.8

[00:18:02]Final eval: perplexity 3.21 — Target convergence: 0.434

[00:18:15]Exporting LoRA weights to isolated artifact store... Done

[00:18:22]Zero retention: ephemeral GPU container flushed

[00:18:30]Serving endpoint online: https://api.langtrain.ai/v1

Listening for cluster events...

Model: mistral-7b-custom-v1

SOVEREIGN TRAINING PIPELINE

From a JSONL file to a live
endpoint. One pipeline.

Full documentation

Stage 03 • Interactive Parameter Studio

Live LoRA Rank & Hyperparameter Modeler

LoRA Rank (r)

α = 128r = 64

r=4 (Minimal)r=16r=32r=64 (Standard)r=128 (Maximum)

Per-Device Micro-Batch Size8 (Grad Accum: 4)

18 (Optimal)1632

Peak Learning Rate (η)2.0e-5 (Cosine)

5e-62e-5 (Default)5e-5

Reactive Telemetry Live Modeler

Trainable Parameters:41.6M (0.06%)

Adapter Weight Size:160 MB

Est. Training VRAM:24 GB (93% savings)

Hardware Feasibility:Single GPU (24GB VRAM)

Fine-tuning a 70B parameter model with LoRA rank 64 updates only 0.06% of total weights, reducing memory overhead from 320 GB to just 24 GB.

PARAMETER-EFFICIENTLive with rank r=64

LoRA & QLoRA

Parameter-efficient fine-tuning with 4-bit and 8-bit quantization. Adapt 70B parameter models on modest GPU clusters with zero precision loss.

VRAM REQUIRED FOR 70B MODEL-93% VRAM

Full 16-bit Fine-tune320 GB

LoRA (Rank 64)52 GB

QLoRA (NF4 Quantized, r=64)24 GB (Single GPU)

Read guide

INSTRUCTION TUNING

Supervised fine-tuning

Full control over instruction tuning, chat formatting, and custom loss masking across conversation turns.

loss_mask: turn_tokens_only

{"role": "user", "content": "..."}

{"role": "assistant", "loss": true}

Read guide

FOUNDATIONAL

Full fine-tuning

Multi-node distributed training with DeepSpeed ZeRO-3 and FSDP for foundational domain adaptation.

DISTRIBUTED: DEEPSPEED ZERO-38x H100 80GB

Architecture specs

ALIGNMENT

RLHF & DPO

Align models with Direct Preference Optimization and reward modeling using pairwise feedback.

PAIRWISE: CHOSEN VS REJECTEDβ = 0.1

Read docs

DATA ISOLATION & SOVEREIGNTY

Trained in isolation. Yours to run wherever you choose.

Single-Tenant Isolated ClusterZERO DATA RETENTION

CUSTOMER ISOLATED VPC BOUNDARYEPHEMERAL RUNTIME

GPU Compute

PyTorch 2.5 + CUDA

Zero Retention

Flushed on job end

OUTBOUND WEIGHTS EXPORTCUSTOMER S3 / HF REGISTRY →

Training happens in single-tenant isolated clusters with zero persistent storage after job completion. Weights belong entirely to you with no platform lock-in.

Security whitepaper

Langtrain Cloud Serverless

Deploy your fine-tuned model to serverless infrastructure with automated scaling down to zero.

https://api.langtrain.ai/v1/models/mistral-7b-custom/completions

Cold start

< 800ms

OpenAI API

100% Compat

P95 Latency

~14ms

View deployment docs

SURFACES

The same pipeline, from
whichever surface suits the job.

Langtrain Studio

Interactive visual workflow editor for dataset curation, hyperparameter sweeps, and lineage inspection.

Explore Studio

Langtrain SDK

Python & TypeScript client libraries to trigger runs directly from your training scripts and notebooks.

Explore SDK

Langtrain CLI

Command-line interface for CI/CD automation, cluster monitoring, and artifact export.

Explore CLI

Langtrain API

REST API with OpenAI-compatible inference endpoints and webhook event subscriptions.

Explore API

01/In code

The same five steps, in Python.

Every snippet is copied from the docs, so what you see is what you run.

training_data.jsonl

from /docs/quick-start

{"messages": [{"role": "user", "content": "What is Langtrain?"}, {"role": "assistant", "content": "It's a platform for fine-tuning LLMs."}]}
{"messages": [{"role": "user", "content": "How do I start?"}, {"role": "assistant", "content": "Use the LoRATrainer class!"}]}

FREQUENT QUESTIONS

Before you start.

Start training.

No GPU reservation required. 50 free compute minutes included.

Start training free Book enterprise demo

Own your weights.Deploy anywhere.

Own your weights.Deploy anywhere.

From a JSONL file to a live
endpoint. One pipeline.

Live LoRA Rank & Hyperparameter Modeler

LoRA & QLoRA

Supervised fine-tuning

Full fine-tuning

RLHF & DPO

Start from the best open models. Leave with your own.

Trained in isolation. Yours to run wherever you choose.

Langtrain Cloud Serverless

The same pipeline, from
whichever surface suits the job.

Langtrain Studio

Langtrain SDK

Langtrain CLI

Langtrain API

The same five steps, in Python.

Before you start.

Start training.

From a JSONL file to a live
endpoint. One pipeline.

Live LoRA Rank & Hyperparameter Modeler

LoRA & QLoRA

Supervised fine-tuning

Full fine-tuning

RLHF & DPO

Start from the best open models. Leave with your own.

Trained in isolation. Yours to run wherever you choose.

Langtrain Cloud Serverless

The same pipeline, from
whichever surface suits the job.

Langtrain Studio

Langtrain SDK

Langtrain CLI

Langtrain API

The same five steps, in Python.

Before you start.

Start training.

Own your weights.Deploy anywhere.

From a JSONL file to a live endpoint. One pipeline.

Live LoRA Rank & Hyperparameter Modeler

LoRA & QLoRA

Supervised fine-tuning

Full fine-tuning

RLHF & DPO

Start from the best open models. Leave with your own.

Trained in isolation. Yours to run wherever you choose.

Langtrain Cloud Serverless

The same pipeline, from whichever surface suits the job.

Langtrain Studio

Langtrain SDK

Langtrain CLI

Langtrain API

The same five steps, in Python.

Before you start.

Do I need ML expertise?

What data format should I use?

How much data do I need?

How long does training take?

Is my data private?

Can I export and self-host my model?

Start training.

From a JSONL file to a live endpoint. One pipeline.

Live LoRA Rank & Hyperparameter Modeler

LoRA & QLoRA

Supervised fine-tuning

Full fine-tuning

RLHF & DPO

Start from the best open models. Leave with your own.

Trained in isolation. Yours to run wherever you choose.

Langtrain Cloud Serverless

The same pipeline, from whichever surface suits the job.

Langtrain Studio

Langtrain SDK

Langtrain CLI

Langtrain API

The same five steps, in Python.

Before you start.

Do I need ML expertise?

What data format should I use?

How much data do I need?

How long does training take?

Is my data private?

Can I export and self-host my model?

Start training.

From a JSONL file to a live
endpoint. One pipeline.

The same pipeline, from
whichever surface suits the job.

From a JSONL file to a live
endpoint. One pipeline.

The same pipeline, from
whichever surface suits the job.