Research
Inside the work.
Experiments and technical notes on models, agents and AI systems. The methods, the results, and what remains to be understood.
- OtherPrompt Optimization on Tasks Nablo Generated ItselfNablo generated a verified evaluation from a demo customer's database and optimized its workspace agent's instructions locally. The pass rate on unseen tasks rose from 22% to 67% with an eight-line instruction edit.4 min
- Small modelsSmall-Model DistillationExperiments in fine-tuning, teacher feedback and model distillation. Each write-up explains the method, the evaluation and its limits.2h 11m
7 parts
- Part 1: Offline Teacher-Trace SFT for a 0.8B SQL Agent25 min
- Part 2: Off-Policy Soft-Label KD for a 0.8B SQL Agent23 min
- Part 3: DAgger-Style Expert-Correction SFT for a 0.8B SQL Agent28 min
- Part 4: On-Policy Probability Distillation for a 0.8B SQL Agent21 min
- Part 5: Training a 0.8B SQL Model with Its Own Feedback12 min
- Part 6: Training a 9B Model to Fix SQL Queries12 min
- Part 7: Training a 4B Model for Customer Support10 min
Your company. Your workspace.
Bring Nablo to your company.
Talk with us about the data your team works with and what you want Nablo to do.