← Back to nablo.io

Research and writing

Technical case studies and post-training experiments behind Nablo's workload-to-model product. The SQL series shares one frozen benchmark, so its numbers compare to each other. The methods that did not work are here too.

Technical case study July 26, 2026 115 / 220

Matching GPT-5.5 medium on one workload with a 9B open model

Post-training a 9B open model on one bounded workload until it reached the same score as GPT-5.5 medium, on a benchmark frozen before the work started.

Method research, 0.8B July 16, 2026 69 / 220

Small-Model Distillation — Part 5: On-Policy Self-Distillation for a 0.8B SQL Agent

No larger teacher at all. The student samples its own actions and trains on probability feedback from a frozen copy of itself.

Method research, 0.8B June 22, 2026 69 / 220

Small-Model Distillation — Part 4: On-Policy Probability Distillation for a 0.8B SQL Agent

A 35B teacher scores the student's own sampled actions with log probabilities instead of writing targets for it.

Method research, 0.8B June 15, 2026 67 / 220

Small-Model Distillation — Part 3: Student-State Hard-Token Correction SFT for a 0.8B SQL Agent

Teacher traces start from the task, so they never cover the mess the student's own policy creates. This one corrects the student at its own failure states, which is the method the 9B case study later scaled up.

Method research, 0.8B June 8, 2026 55 / 220

Small-Model Distillation — Part 2: Off-Policy Soft-Label KD for a 0.8B SQL Agent

Same student, same harness, one change. The student sees the top-k distribution behind each teacher choice instead of only the final pick.

Method research, 0.8B May 30, 2026 46 / 220

Small-Model Distillation — Part 1: Offline Teacher-Trace SFT for a 0.8B SQL Agent

A 0.8B model copies successful teacher trajectories one token at a time. It went from 1 solved task to 46, which taught it to operate the agent loop but not to write better SQL.