Browse categories
Explore
Fiverr Pro
English
$
USD
I will fine tune a llm or slm based one the requirements
Eshan E
About this gig
I fine-tune small open models and ship the evaluation that proves it worked or tells you honestly that it didn't.
WHAT YOU GET
LoRA/QLoRA fine-tune on Llama 3.1, Qwen 2.5, Mistral or Phi
Dataset cleaning and formatting
Held-out evaluation against your baseline, with real numbers
Quantized GGUF weights you can run locally
All training code and configs
A short report: what changed and what I'd do next
WHY ME
First-author paper under review at ICLR 2027 on programmatic LLM evaluation
Shipped a production LLM assistant live to 70+ training centres
Built an AI agent with 15+ native tools and strict JSON dispatch
HOW IT WORKS
Message me before ordering with your dataset and what the model should do. I'll tell you if fine-tuning is the right answer sometimes it isn't, and I'd rather say so than take your money.
I build the eval set first and run your baseline, so "better" is a number, not an opinion. If it won't beat your baseline, you hear it in days, not at delivery.
Your data is used only for your project and never shared.
Get to know Eshan E
Eshan E
Forward Deployed Engineer
- FromIndia
- Member sinceSep 2026
- Avg. response time1 hour
Languages
English, Hindi
I am a Forward Deployed Engineer and AI Developer with expertise in building confidential document de-identification pipelines and production-grade LLM assistant systems. I specialize in Python, PyTorch, and Kubernetes, with a strong focus on AI observability, model evaluation, and self-healing cloud infrastructure. I have a proven track record in competitive programming and research in transformer language models.
My Portfolio
FAQ
What do you need from me to start?
Your dataset (JSONL, CSV or plain text), a sentence on what the model should do, and how you're currently doing it — a prompt, an API call, or an existing model. That last part matters: it's the baseline I measure against.
What if the fine-tune doesn't beat my current setup?
You hear it in the first few days, with the numbers. Sometimes better prompting or retrieval is the real answer and fine-tuning won't help. I'd rather tell you that early than deliver something that doesn't work.
How much data do I need?
A few hundred good examples often beats tens of thousands of noisy ones. Basic covers up to 5,000 rows. If you have less, message me — it's usually still workable. If you have far more, Standard includes the cleaning.
Can I run the model myself afterwards?
Yes. Every package ships the merged weights and the training code. Standard and Premium add a quantized GGUF that runs locally on CPU or a consumer GPU. Premium adds a FastAPI service you can deploy.
Is my data safe?
Your data is used only for your project, never shared, and never used to train anything else. I currently build a de-identification pipeline for confidential client documents, so this is the discipline I work in daily.
