Training
Each walkthrough takes a real workload (SQL generation, contract review, support classification) and uses Blueprint to ship a custom-calibrated, fine-tuned, or compressed model that beats off-the-shelf on the eval set. Video, transcript, and a downloadable sample corpus included.
Route 75% of traffic to a 3B model and 25% to a 70B, with measured per-class accuracy. End result: ~5× lower average inference cost than 70B-everywhere, with quality on hard queries preserved.
A 3.8k-token system prompt squeezed to 850 tokens via LLMLingua-2, with measured accuracy retention on a held-out QA set. End-to-end through Blueprint's Compress card.
End-to-end QLoRA run on Blueprint: 4-bit base, rank-16 adapter on attention + MLP, 3 epochs on 3000 contract clauses. The result is a 7B that beats the off-the-shelf 70B on the trained task at a fraction of the inference cost.
End-to-end: take a customer-support intent classification workload, calibrate Q4 against actual support transcripts, ship a custom GGUF that matches the f16 base on accuracy at 3× lower memory.
End-to-end walkthrough: from production SQL prompts to a custom-calibrated Q4_K_M GGUF that beats bartowski's stock quant by 9 percentage points on a held-out SQL benchmark.