LLM Fine-Tuning/post training | 4-8 Week Open-Weight Model Adaptation

We're hiring a senior engineer for a focused proof of concept. Full details shared with shortlisted candidates. THE PROJECT Prove, within 4-8 weeks, that an open-weight base model adapted with a curated sample of our proprietary data produces measurable improvement on culturally specific tasks: local references, language nuance, culturally appropriate responses instead of generic default answers. At the final review, stakeholders run agreed prompts against the base model and the adapted model side by side and see where our data improved the output. This is a feasibility sprint, not production work. Potential follow-on work if it goes well. DELIVERABLES The full list below is due within the 4-week window, with a progress review at the midpoint and a final review at the end. Sequencing is yours to propose; you are the expert. What is fixed is the complete list: - Written base model selection and benchmarking recommendation against our tasks and data-sovereignty constraints. - A controlled, reproducible environment for inference and adaptation (versioned configs, run tracking). - Curated data sample and a clean held-out evaluation set. - Evaluation framework v0 that scores base vs adapted outputs across agreed domains, with baseline results established before adaptation. - Adaptation experiment (LoRA/QLoRA/SFT/distillation as the data dictates and licenses permit) with a quantified before-and-after against the held-out set. - A single-language multilingual proof of concept. - Demo-ready V0 supporting the side-by-side comparison on agreed prompts. - Initial governance, safety, and provenance framework. - TPU/GPU deployment assessment. - Integration spec for our agent and routing layer. Phase 2 roadmap. - Full handover package: source code, configs, adaptation scripts, prepared datasets, evaluation and architecture documentation. CONSTRAINTS - All work runs inside our controlled environment. No company data touches any third-party or externally hosted inference service. Compute is provided; you tell us what to provision. Open-weight licenses must permit commercial use. OUT OF SCOPE Production deployment and MLOps, full-scale training runs, ingestion of the full corpus, broad domain and language coverage, and any guaranteed quality numbers. We report what the evals show. REQUIREMENTS - Hands-on experience fine-tuning open-weight LLMs (not just API fine-tuning), building evaluation harnesses with held-out sets, and working with data that cannot leave a controlled environment. - Experience with multilingual or non-English model evaluation is a strong plus. You personally lead and stay hands-on; a small support team is fine but the core work is not delegated. TO APPLY, answer these directly (applications that skip them are ignored): - Describe the largest open-weight fine-tuning project you have personally executed: model, technique, dataset size, and what measurable improvement you achieved. - How would you build an evaluation framework for subjective cultural quality, where "better" is not a benchmark score? Be specific about held-out sets, judges, and gold standards. - Have you deployed or assessed deployment on TPUs as well as GPUs? Describe. - Given the full deliverables list above, how would you sequence the 4 weeks? Brief outline is fine; we want your plan, not ours. Can you start soon, and can you commit the full 4-weeks without competing obligations?