new·The score now tells you which way it movedA brain's exam only ever grows: its own material writes questions, and so does every question a real caller asked and did not get answered. The score is a percentage over that growing set, so a brain that learned more could post a smaller number — and this week three did. One of them answered two MORE questions than the week before and showed eighteen points less. Printed as a single percentage, that reads as decline to a reader and as punishment to anyone who contributes material.all news →
mozg.beta
Sign in

Building With AI Agents · all subjects

deploy/models

2 notes, read out of this brain and free to use. Each one was extracted from a source and is re-checked against its exam.

Model weight files must go to their target-specific destinations, not environment variables

Setting DEST=/models/bge-m3 in an image meant fetch-model.sh honoured it for every call, so `fetch-model.sh reranker` wrote the cross-encoder's safetensors into the embedding model's folder. Transformers prefers safetensors, so the next restart would have loaded a sequence-classification model as the embedder. The destination must come from the target, with an optional second argument for a custom path, and the image passes its path to the boot fetch instead of leaving it in the environment.

Missing or broken model weights fail loudly, not silently

bge-reranker-v2-m3 ships model.safetensors but the fetcher asked for pytorch_model.bin and got a 404. The directory already held an embedding model's files, so transformers loaded a checkpoint with no classifier head, built a random one, and only warned. The service came up healthy but reordered every search by noise. This is strictly worse than having no reranker. The model must sit one obvious pair before it will serve anything: if the relevant document does not beat the unrelated one, loading fails loudly and search falls back to RRF order.

Give your agent this brain