Frequently Asked Questions

LangTrain supports 50+ open-source models including Llama 3.x, Mistral, Gemma, Phi 4, Qwen 2.5, DeepSeek, Falcon, and more. You can also bring your own model checkpoints from HuggingFace.

Fine-tuning time depends on model size, dataset size, and method:

- QLoRA 7B model: 30-60 minutes
- LoRA 13B model: 1-2 hours
- Full fine-tuning 70B: 6-24 hours

We use H100 GPUs for maximum performance.

Yes! You can upload your own datasets in JSONL, CSV, or Parquet format. We support:

- Instruction-following format
- Chat format (messages)
- Completion format (text only)

Absolutely. We use:

- End-to-end encryption for all data
- SOC 2 Type II compliance
- Data isolation per workspace
- No training on your data for other models
- GDPR compliant data handling

LoRA (Low-Rank Adaptation):
- Trains only 0.1-1% of parameters
- 10x faster, 10x less memory
- Great for most use cases

Full Fine-tuning:
- Updates all parameters
- Maximum performance
- Requires more compute

Yes! You can:

- Export models to HuggingFace format
- Deploy to AWS, GCP, or Azure
- Use our containerized inference servers
- Self-host with Docker

Yes, we provide:

- REST API for all features
- Python SDK (pip install langtrain-ai)
- OpenAI-compatible inference endpoints
- Comprehensive API documentation

We offer flexible pricing:

- Starter: Free tier for experimentation
- Pro: $49/month with more compute
- Enterprise: Custom pricing for large teams

All plans include pay-as-you-go GPU compute.