How a Dedup Pass Deleted My Training Curriculum
Code: Megapixel99/capture-the-flag I built a capture-the-flag arena where language models attack and defend each other's containers. The replays then became training data for a small local model that plays in the same tournament. The bot is Qwen2.5-3B-Instruct with an MLX LoRA adapter (rank 8, 8 layers) trained on turns extracted from game logs. It does well on offense. Across the 221 logged…
A language model called Qwen2.5-3B-Instruct with a LoRA adapter was trained on captured flags from a capture-the-flag tournament. The model performed well in offense but struggled in defense. Two attempts to improve defensive performance were rolled back due to regressions. The retraining process involved weighting defensive turns, but the weighting stage and deduplication stage were the same operation with opposite effects, resulting in the weights being discarded during deduplication.
Only turns scored 0 (incorrect commands) survived, leading to a binary include-or-exclude filter, effectively negating the intended curriculum.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
