AI Post-Training, Fine-Tuning & Model Alignment
A generic model will not know your contracts or your output format until it is trained on real examples of both. We run post-training as disciplined engineering, on data you approve, with evals agreed before training and everything inside your boundary when you need that.
An eval set first, so better is measured, not claimed
Fine-tuning fails quietly when no one can say what better looks like. We agree a benchmark and a regression suite before the first run, then report every run against it.
What this looks like in the codebase
After the workflow is clear, we lock the technical choices that keep the product maintainable. These are the ones we use here:
- Eval-first: a domain benchmark and regression suite agreed with you before any training run
- Auditable data pipeline: curation, dedup, quality filters, and synthetic augmentation where real data is thin
- Supervised fine-tuning first; RLHF or DPO preference tuning when tone and behavior have to be right
- LoRA and PEFT for fast, cheap iteration, with a full fine-tune path when the adapter hits its ceiling
- Private training: inside your VPC and accounts, or under vendor terms with no training retention
Stack we ship with
Chosen because we have run it in production, not because it is fashionable.
The work inside this service
From the first data model to a production deploy. Here is what we hand over.
Domain Fine-Tuning & PEFT
SFT on curated examples from your documents, tickets and transcripts. LoRA iteration while we learn, a full fine-tune only when the data says it pays.
- Supervised fine-tuning on your domain data with format and style controls
- LoRA and QLoRA runs for cheap, fast iteration
- Full fine-tune path when an adapter cannot carry the behavior
- In-VPC training and no-retention variants for sensitive data
Alignment & Preference Tuning
RLHF and DPO for behavior that examples alone cannot fix: tone, refusals and format discipline. Guardrails checked before anything ships.
- RLHF and DPO preference tuning for tone and judgment
- Instruction following and strict output-format adherence
- Safety and guardrail evals on adversarial and edge inputs
- Red-team style checks where the stakes justify them
Evals & Model Lifecycle
Before and after numbers on your eval set, and a record of every trained version. A handoff to serving that does not come undone at deployment.
- Domain eval sets and regression harnesses agreed before training
- Before/after reporting including cost per point of improvement
- Model registry and versioning across every run and dataset
- Handoff to serving (vLLM and your stack) or a hosted API
A call we made in production
Legal-ops summarization, fixed with a LoRA adapter
Legal-ops summarization, fixed with a LoRA adapter
For a legal-ops summarization product, the general model drafted fine but mangled clause citations and ignored the firm's required output headings. We built a regression set of 400 judged summaries plus a 150-case benchmark before touching any weights. Training data came from 12,000 approved contracts and memos, deduplicated and filtered down to 9,400 quality examples, then augmented with synthetic summaries for formats the real data under-covered. A dozen LoRA runs across two H100s cost about $850 total, each finished in hours, and every run was reported against the same eval set. Citation accuracy went from 71% to 94%, and reviewers stopped flagging the output format. A full fine-tune would have lifted accuracy one more point for six times the spend, so the adapter shipped.
Questions before you write to us
Ready to turn this into a product? Tell us what you need.