Frequently Asked Questions

Langtrain is a developer platform for fine-tuning and deploying Large Language Models. You can customize models like Llama 3.3, Mistral, Qwen, and DeepSeek with your own data using a simple Python API. Train a 7B model on 1,000 examples in about 20 minutes with just a few lines of code.

Fine-tuning gives your model deep understanding of your data, not just surface-level instructions:

• –Larger Context: Your model internalizes patterns from thousands of examples, far beyond what fits in a prompt
• –Better Understanding: The model learns how to respond in your domain, not just what to respond
• –Personal LLM: Your fine-tuned model becomes your personalized AI that understands your business, writing style, and requirements
• –Automation Ready: With agents, your personal LLM can automate workflows because it truly understands your context

Prompts are great for testing. Fine-tuning is how you build production-grade AI that works.

Absolutely! This is one of the most powerful use cases:

• –Context-Aware Automation: Your fine-tuned model understands your business domain, making agent decisions more accurate
• –Custom Workflows: Build agents that draft emails in your voice, analyze data your way, or handle customer queries with your expertise
• –Langtrain Agents: Deploy your model as an agent with tools (web search, APIs, databases) directly from our platform
• –External Integration: Export to any agent framework (LangChain, AutoGen, CrewAI)

Your personal LLM + agents = powerful automation that actually understands what you need.

No. Langtrain handles the complexity of model training—you just provide your data. Our SDK abstracts hyperparameter tuning, distributed training, and optimization. If you can write Python, you can train models.

We support 100+ open-source models:

• –Llama 3.3: 8B, 70B, 405B
• –Mistral/Mixtral: 7B, 8x7B, 8x22B
• –Qwen 2.5: 7B, 14B, 32B, 72B
• –DeepSeek V3: 70B, 236B
• –Gemma 2: 9B, 27B
• –Phi-4: 14B

You can also import any model from HuggingFace.

We accept JSONL, CSV, and Parquet files. The recommended format is JSONL with chat messages:``json{"messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}`

We also support instruction format (instruction, input, output`) and raw text completion.

Quality matters more than quantity:

• –50-100 examples: Good for style/tone adaptation
• –500-1K examples: Recommended for task-specific training
• –10K+ examples: Great for domain expertise

We provide data preprocessing and augmentation tools to help maximize your dataset's effectiveness.

Depends on model size and dataset:

• –7B model, 1K examples: ~20 minutes
• –13B model, 5K examples: ~1 hour
• –70B model, 50K examples: ~8 hours

We use NVIDIA H100 GPUs. You can monitor progress in real-time via dashboard or webhooks.

LoRA/QLoRA trains adapters (~1% of parameters). It's 10x faster, uses 10x less memory, and works great for most tasks. Full fine-tuning updates all parameters for maximum performance but requires more compute. We recommend starting with LoRA—most users never need full fine-tuning.

Yes. Your data is:

• –Encrypted at rest (AES-256) and in transit (TLS 1.3)
• –Isolated in your organization—never shared
• –Never used to train other models
• –Deletable on request (GDPR compliant)

We are not yet SOC 2 certified; an audit has not been completed. For healthcare or other regulated data, the self-hosted deployment keeps data inside your own environment.

Absolutely. Export to:

• –HuggingFace format for any inference framework
• –GGUF for llama.cpp and local inference
• –Docker container with our optimized server
• –AWS/GCP/Azure with one-click deployment

You own your trained models completely.

There are three plans:

• –Starter: free, 5M training tokens
• –Pro: $20/month (₹1,499 in India), 200M training tokens and priority GPU access
• –Enterprise: custom pricing, private deployment

Yearly billing is 20% less. See Pricing for the full comparison.

Install the SDK and train your first model:``bashpip install langtrain-ai`

Then in Python:

`pythonfrom langtrain import LoRATrainertrainer = LoRATrainer(model="llama-3.3-8b")trainer.train(data)``See our Quick Start guide for a complete walkthrough.

Multiple support channels:

• –Documentation: Comprehensive guides at langtrain.xyz/docs
• –Discord: Active community with 5K+ developers
• –GitHub: Open-source examples and issue tracking
• –Email: support@langtrain.xyz
• –Enterprise: Dedicated Slack channel + SLA