tinyforge teaches tiny models from failures

The most interesting result here is not the raw benchmark bump but the claim that tiny models can learn how to use verifier feedback, not just memorize answers. If that holds up, tinyforge points to a cheap local recipe for self-improving systems in any domain with automatic checks.

• –The project combines evolutionary search, exact test-failure feedback, and LoRA fine-tuning instead of relying on a larger teacher model
• –The repo argues the biggest gains come from the repair loop itself, with the trained adapter becoming a better “repair partner” when shown what failed
• –Resource requirements are unusually approachable for this kind of work: Apple Silicon, 6GB-13GB peak memory, and a few minutes of training
• –The code is MIT-licensed and already packaged as a runnable CLI, which makes it more than a one-off Reddit experiment
• –The limitation is explicit: this does not magically turn a 0.8B model into GPT-4 class performance; the gains are scoped and system-dependent