13-stage PyTorch checkpoint set for Parselex, a deterministic, offline resume-to-JSON extraction pipeline. No LLM involved — each stage is a small MiniLM-backbone classifier (~23M params) fine-tuned for one step of the pipeline (section detection, entry boundary detection, field classification, run separately per resume section).

This repo hosts weights only — no inference code. Clone the code repo and point it at these weights: https://github.com/karan-963/parselex

Both FP32 and INT8 (quantized) checkpoints for all 13 stages, ~2.5GB total:

best_model.pt = FP32 checkpoint. best_model_int8.pt = quantized (smaller, faster, slightly lower accuracy). Folder layout matches model_weights/<stage>/ in the code repo exactly — no renaming needed to use these.

Or manually:

See the paper for full methodology, held-out test numbers, and known limitations (small training set, single-column resumes only, synthetic-data ceiling): <arXiv link — filled in once submitted>

MIT — same as the code repository. See LICENSE.