mjay1337/needle-tag-search

Needle Tag Search

Needle checkpoint fine-tuned to turn natural-language imageboard searches into candidate-constrained postfix Boolean plans. EroHaven combines it with a deterministic grammar and resolves the resulting operands against canonical tags supplied by BooruDex.

The repository contains:

• model.pkl: accepted v6 CPU fallback checkpoint.
• needle/: patched Needle runtime at upstream commit ffb1c5144c5a16cb8ec650dbc8a6f6fd3854f8f2, including static encoder padding to avoid recompiling JAX for each query length and a runtime-only tokenizer import that avoids loading the training dataset stack.
• LICENSE: upstream Needle MIT license.

The checkpoint SHA-256 is 3f2afc521eede4bd6972bb462de66248d3e36ed384c5fe796a9416b9a3fc99f1.

This model handles tags from imageboards across all rating categories and may emit explicit tag names. It is a query parser, not a content-safety classifier. The checkpoint uses Python pickle and should only be loaded from a trusted, pinned repository revision.