Frequently Asked Questions

Langtrain fine-tunes open models such as Llama, Mistral, Qwen and Gemma on your own data, then gives you the weights. Train on your own GPU with the open-source Python libraries, or in the Langtrain cloud from the dashboard or the API.

A prompt can only show a model a few examples. Fine-tuning teaches it from hundreds or thousands, so it learns your format, tone and domain and follows them without a long prompt every time. Start with prompting; fine-tune when prompts stop being enough or get too long and slow.

No. If you can write a JSONL file and a few lines of Python, you can train a model. The defaults (QLoRA, rank 16, three epochs) are a sound starting point, and the dashboard needs no code at all.

Open text models from Hugging Face that Transformers and PEFT support, such as Llama, Mistral, Qwen, Gemma and Phi. Gated models like Llama need you to accept their licence on Hugging Face first. For images, Langvision fine-tunes vision-language models; see Fine-tune vision models.

JSONL with one conversation per line, in the OpenAI chat format:``json{"messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}``For preference tuning, each line has prompt, chosen and rejected. See Datasets.

Quality matters more than quantity. A few hundred clean, consistent examples can change a model's tone and format; teaching a narrow task well usually takes more. Remove duplicates and contradictory answers first: DatasetIntelligence in the Python SDK reports both.

QLoRA is the default: it loads the base model in 4 bits and trains a small adapter, so 7B and 8B models fit on a single 16 GB GPU. LoRA does the same on a 16-bit model and needs more memory. DoRA, IA³ and prefix tuning are other adapter methods. See LoRA & QLoRA.

Train with Langtune on your own hardware and your data never leaves it. In the Langtrain cloud, every connection uses TLS and stored data is encrypted at rest by our database provider. We are not yet SOC 2 certified; an audit has not been completed. See Security.

Yes, the weights are yours. Merge the adapter into the base model to get a standard Hugging Face folder, convert it to GGUF for llama.cpp and Ollama, serve it with vLLM, or push it to your own Hugging Face repo. See Run your model anywhere.

Training on your own GPU with the open-source libraries is free. For the cloud:

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

Yearly billing costs less. See Pricing.

Follow the Quick start: install the SDK, write a small JSONL file and train on your GPU.

• –Docs: you are here
• –GitHub: issues on langtune and langvision
• –Email: support@langtrain.xyz

See Support.