Qwen3.5-9B · factorio-build · Evolve & Reinforce round 1 warm start (SFT)
The SFT warm start for qwen3.5-9b-factorio-build-er-r1. Base Qwen3.5-9B was shown 800 evolved builder programs and asked to explain and write each one itself. The 279 replies whose programs still passed were kept and trained on with the program removed from the prompt, plus 20% replay of the base model's own answers to general prompts. LoRA rank 32, 3 epochs, lr 1e-4, merged. Thinking off, factorio-build prompt v2.
| Base | This model | |
|---|---|---|
| Factorio val success | 2.3% | 22.3% |
| PlanBench plan generation, first 100, thinking off | 53% | 49% |
| Reply tokens | 1,975 | 1,859 |