Private language models, deployed inside your own network. ExperimentLab · Varanasi, India
I fine-tune small language models for one domain, evaluate them against your own documents, and install them behind your firewall — together with the website, API and automation built around them. No prompt or document leaves your network, and there is no per-token bill.
Five Apache-2.0 LoRA adapters, with a Colab chat demo for each. Adapters are weights only — apply them to the base model named in the row.
| Model | Domain | Base model | Adapter | Try |
|---|---|---|---|---|
| Vidhi-AI-Instruct | Indian legal reasoning, statutory compliance, contract risk | Qwen2.5-7B-Instruct, r=8 | 10 MB | Colab |
| olmo2-7b-silicon-rtl-eda | Verilog HDL, RTL synthesis, timing closure | OLMo-2-1124-7B-Instruct, r=16 | 160 MB | Colab |
| olmo2-7b-phd-pure-math | Proof synthesis, algebra, differential geometry | OLMo-2-1124-7B-Instruct, r=16 | 160 MB | Colab |
| olmo2-7b-biomed-chem | Molecular informatics, organic synthesis, pathways | OLMo-2-1124-7B-Instruct, r=16 | 160 MB | Colab |
| olmo2-7b-astro-logic | Orbital dynamics, stellar mechanics, relativistic calculation | OLMo-2-1124-7B-Instruct, r=16 | 160 MB | Colab |
Each Colab loads the base model plus adapter in 4-bit NF4 on a free T4 and opens a Gradio chat — it is a hands-on demo, not a benchmark harness.
Corpora I built and published: Hinglish-English STEM 500k · 1B STEM pretrain set · Vidhi-AI 1k curated
Backend. Product APIs, admin panels and data stores deployed as a single self-hosted binary on a VPS you own: schema and row permissions, authentication, file storage, realtime updates and server-side hooks. You get the repository, the server and the backup schedule — not a vendor lock-in.
Web frontends. TypeScript and React product UI, shipped on static hosting. Public example: craftora — 28 production dependencies, IndexedDB persistence, and no server code anywhere in the repository.
Model integration. The adapters above, or your model selection, behind your own endpoint: retrieval over your documents, streamed responses, per-key rate limits, and an evaluation set that stays with you.
Workflow automation. WhatsApp, Google Sheets, PDF and Tally or ERP pipelines. The model handles the judgement step — classify, extract, draft — and deterministic code handles the transaction. Every rule is readable, because a business cannot audit a prompt.
- Readiness audit — 2 to 3 weeks, fixed fee. We benchmark candidate models on a sample of your documents and give you a written go/no-go: accuracy ceiling, failure modes, hardware sizing, cost of ownership. If a small model will not do the job, the report says so.
- Pilot — fine-tuning on your corpus, scored against the same held-out set, so improvement is measured rather than asserted.
- Deployment — weights, adapters and vector store on hardware you own, on your network, with a documented rebuild path. Retainer covers refresh, re-evaluation and uptime.
Stack: Unsloth, Hugging Face TRL, PEFT/LoRA, NF4 quantization, vLLM, Docker, Qdrant. Product side: TypeScript, React, Vite, IndexedDB, static and edge hosting, self-hosted single-binary backends.
- These are 7B-class models. They win on privacy, latency and cost — not on open-ended reasoning. An audit tells you which one you need before you buy hardware.
- Air-gapped deployment supports your DPDP or ISO 27001 obligations. It does not discharge them.
- You deal with the engineer who builds and deploys it. No account managers, no bench.
Every model, dataset and demo above links to something you can open right now. If I cannot link it, I do not claim it. Client work is described under NDA where the client asks for it; references come before you commit to anything.
Email: shreyansh@experimentlab.in Hugging Face: shreyansh12183