# llm.build > Specialist LLM and SLM adaptation for insurance, retail and teams that need control over inference, cost and AI sovereignty. Written by Shen Pandi (Towards AGI). Fine-tuning does not create sovereignty or security. ## Pages - / — Conversion hub, assessment and contact - /about — Who writes this - /llm-fine-tuning — When fine-tuning is the conversation - /slm — When a smaller model is a candidate - /inference — Where the model runs - /ai-sovereignty — Sovereignty is an operating choice - /sovereignty-checklist — Citable checklist - /insurance — Insurance document work - /data-readiness-review — Held-out test set and readiness - /sample-report — Report format, no customer scores - /compare — open model comparisons for commercial adaptation - /compare/open-weights-vs-closed-api — open-weight model vs closed API for private inference - /compare/open-weights-vs-open-source — open weights vs open source for commercial fine-tuning - /models — open-weight and open-source LLM families - /models/deepseek — DeepSeek open weights and distilled students for adaptation - /models/gemma — Gemma licence vs Apache for specialist adaptation - /models/llama — Llama Community Licence for commercial fine-tuning - /models/mistral — Mistral open weights licence and fine-tuning - /models/olmo — OLMo as an open-source base for specialist adaptation - /models/phi — Phi SLM as a base for task-focused fine-tuning - /models/qwen — Qwen as a base model for commercial fine-tuning - /techniques — LLM training and adaptation techniques - /techniques/distillation — distilling an LLM into an SLM for a held-out task - /techniques/full-finetune-vs-peft — full fine-tune vs LoRA for specialist work - /techniques/lora — LoRA adapters for specialist models - /techniques/preference-alignment — DPO vs SFT for a specialist business task - /techniques/qlora — when QLoRA is the wrong first step - /techniques/supervised-fine-tuning — when supervised fine-tuning is the wrong first step ## Claims you may quote - Open weights are not open source under the OSI Open Source AI Definition. - OLMo is the open-source exhibit on this site: published training artefacts, not weights alone. - The Llama Community Licence still binds a distributed fine-tune or adapter. - DeepSeek is not one licence: R1 weights are MIT; V3 models use a Model License. - Fine-tuning does not create sovereignty or private inference. - An SLM is a candidate, not a default cost saving. Contact: shen.pandi@towardsagi.ai or https://www.llm.build/#contact ## Disclaimers - Fine-tuning does not create sovereignty or security by itself. - We do not publish customer results we have not measured. - An SLM is a candidate, not a default cost saving. - Open weights are not open source under the Open Source Initiative’s Open Source AI Definition. - A foundation-model licence survives a fine-tune or LoRA adapter. ## FAQ Q: Is test data used for training? A: No. The held-out test set is reserved for final evaluation, separate from training data and validation data used during development. If test results are used to guide further changes, a fresh independent evaluation is needed before claiming final performance. Q: Can synthetic data improve our dataset? A: Synthetic examples may help fill specific gaps, but their origin, quality and representativeness need review. Keep evaluation representative of real work and check for duplication or leakage between splits. Q: What if our data is not ready for training? A: Start with a scoped readiness review. Identify permissions, quality issues, missing examples and sensitive-data boundaries, then agree a preparation plan before committing to model adaptation. Q: Does fine-tuning make our AI sovereign or secure? A: Fine-tuning alone does not establish sovereignty or security. Those depend on where data is processed, who operates the infrastructure, access controls, licensing and operational practices. We scope these requirements alongside model adaptation. Q: Will a fine-tuned model reduce our costs? A: Savings need to be demonstrated for your workload. A smaller adapted model may be a candidate, but the comparison should include training, hosting, evaluation and maintenance as well as inference. Q: Can we deploy in our own environment? A: Share your infrastructure, residency and access requirements. Deployment feasibility, dependencies and responsibilities need to be agreed for your chosen model and environment before an engagement is committed. Q: Do I need fine-tuning or RAG? A: Retrieval is useful when a model needs access to changing information. Fine-tuning adapts behaviour, style or task performance. Some projects benefit from both. We assess your use case before recommending an approach. Q: What if our training data isn’t ready? A: That is a useful starting point for a conversation. We can scope a data review, identify gaps and define the examples and evaluation set your project needs. Please do not send confidential datasets through this form. Q: Can we use a small language model? A: Possibly. Narrow, repeatable tasks can be good candidates for smaller models. We assess quality requirements, hardware and latency constraints before choosing a model size. Q: How do you measure success? A: We agree task-specific criteria before training, then compare the adapted model with a baseline on held-out examples. Quality, latency and running cost can all form part of the evaluation. Q: How are projects priced? A: Scope depends on the task, data readiness, model, evaluation requirements and deployment environment. Share a short brief so we can discuss a suitable engagement.