Engagement
What’s included in this service?
Base model selection
We choose the base family and size for each task, typically 8B–32B parameters, from Llama, Qwen, Mistral, Gemma and DeepSeek, based on measured accuracy, latency and serving cost.
Supervised fine-tuning and LoRA
Full supervised fine-tuning or LoRA/QLoRA adapters, trained on verified examples captured by BackContext. Most tasks need a few thousand.
Preference tuning
We train on what your team accepted versus what it rejected, so the model learns your conventions instead of generic best practice.
Distillation from frontier models
Where your own examples are thin, we distill from frontier outputs your team has reviewed, then train past them on your outcomes.
