Model training

Generic models
don’t know your
CBK circulars.

When RAG isn’t enough, train adapters on your approved corpus so the model speaks banking regulation, insurance clauses, and Kenyan business English natively.

Off-the-shelf models handle generic English fine. They stumble on CBK prudential guidelines shorthand, Kenyan insurance policy wording, SACCO bylaws, and the tone your compliance team expects in customer letters.

Fine-tuning at Corvux means LoRA/QLoRA adapters on open weights, trained on your curated examples, evaluated on held-out sets, deployed only when metrics clear gates you agree upfront.

Training pipeline

Curate → train → eval → promote.

01
Corpus curation
Approved examples only, redact PII, dedupe, label task types with your SMEs.
02
Baseline eval
Score base model on your test set before training spend.
03
Train adapters
Kenya GPUs when the corpus must stay, or a cloud GPU you approve when it can travel. Managed, on-prem, or cloud hosting after the eval.
04
Red team
Eval suite for hallucination, leakage, and tone before prod.
$ corvux finetune run --project cbk-memo-v1 \
 --base llama-70b-ke --method qlora

CORPUS your credit memos · SME validated
TRAIN kenya-gpu-example · 6h · adapter 142MB
EVAL held-out set you approve, scores after the run
COMPARE base+RAG vs adapter+RAG, scored on your set
✓ staged for prod · rollback: cbk-memo-v0
LoRA
Efficient
Your data
Your adapter
Eval
Gate required
KE
Training

Domain packs we’ve shaped

Kenyan verticals that benefit.

Banking & CBK

Credit memos, regulatory summaries, customer decline letters, banking.

Insurance wording

Claims correspondence, policy endorsements, pairs with sector packs.

Swahili business

Formal customer comms, combine with Swahili pack.

Legal diligence

Clause extraction tone for legal suite outputs.

Fine-tuning without eval is cosplay. We won’t promote an adapter that beats base model on vanity loss but fails your SME rubric.

Serve tuned models via model routing, task A gets base+RAG, task B gets your adapter automatically. A model built for one narrow task starts on custom models. Serving the adapter on a cloud you pick is model hosting.

FAQ

ML and risk stakeholders.

Fine-tune vs RAG, when to choose?

RAG for changing docs. Fine-tune when tone, terminology, and format consistency count across thousands of similar outputs, claims letters, credit memos, regulatory summaries.

How much data do we need?

Hundreds of quality examples for LoRA adapters; thousands for deeper runs. We audit your corpus before quoting, garbage in is still garbage out.

Can we undo a bad fine-tune?

Versioned adapters. Roll back to prior checkpoint; A/B in staging before production promotion.

Who owns the resulting weights?

You do, contractually. We train on your infra or managed tenant; no shared global model trained on your data.

Scope a fine-tune on one task

Bring 50 example inputs/outputs. We tell you honestly if RAG is enough or training will move the needle.

Book a fine-tuning assessment