Browse categories
Explore
Fiverr Pro
English
$
USD
I will fine tune llama, qwen or mistral llms on your data with lora
Sidharth P
About this gig
Need an open-source LLM that really knows your domain, format or style? I will fine tune it on your data and show you, with numbers, how much better it got.
I am an AI researcher (ICML 2026, IJCNLP-AACL 2025) and I fine tune and evaluate open models on multi-GPU H100 setups as part of my research.
What you get:
- Base model and method recommendation (LoRA, QLoRA, full SFT or GRPO)
- Data cleaning and formatting into chat or instruction templates
- Clean, reproducible training code and config
- Trained LoRA adapter, or merged weights on higher packages
- Before and after evaluation on a held-out set
- Optional fast serving setup with vLLM or SGLang
Models: Llama, Qwen, Mistral, Gemma, DeepSeek and other Hugging Face models.
Message me your dataset and goal before ordering and I will suggest the right base model, method and package.
Get to know Sidharth P
Sidharth P
AI Safety Researcher and LLM Fine Tuning Expert
- FromIndia
- Member sinceApr 2017
Languages
Telugu, English, Hindi
AI researcher in LLM safety and NLP. Research Intern at AI4Bharat (first author of IndicBERT-v3, open multilingual encoder LLMs) and Research Fellow at SPAR. Papers at ICML 2026 and IJCNLP-AACL 2025, plus arXiv work on memory attacks against LLM agents. Hands-on with LoRA/QLoRA and full fine-tuning (SFT, GRPO), multi-GPU training on H100s, vLLM/SGLang inference and LLM-as-judge evaluation. I help teams fine-tune open-source LLMs, build evaluation pipelines, red-team LLM agents and reproduce ML papers. Message me your goal and I will reply with a clear plan.
My Portfolio
Other AI Development Services I Offer
FAQ
What data format do you need?
JSONL, CSV or Parquet with instruction and response pairs (or prompt and completion). If your data is raw or messy, I can help clean and format it first.
Do I need my own GPUs?
No. I can run training on cloud GPUs, with compute costs agreed upfront, or train on your own cloud account or hardware if your data must stay with you.
How much data do I need?
For style or format tasks a few hundred good examples can be enough. For new domain knowledge or harder tasks, a few thousand is better. If you have less, I can help you build or augment a dataset.
Will I own the fine-tuned model?
Yes. You get the adapter or merged weights, the training code and the config. Your data and model stay private and are not reused for anyone else. Usage is subject only to the base model's own license.
