ProxyType-4B

ProxyType-4B extracts a structured voting record from historical SEC Form N-PX text and HTML. Each input contains one manually marked voting target and its surrounding source context. The model belongs to the ProxyBench research project. The adapter is available from this Hugging Face repository.

Model and training

An adapter stores learned weight changes separately from its base model. This repository contains a LoRA adapter, which trains small weight changes while keeping the base fixed. The required base is Qwen/Qwen3.5-4B at revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a. The base weights and merged model are separate inputs and are not included here.

Training used 330 accepted examples and completed 660 updates over two epochs. An epoch is one pass through the training examples. The dataset contains 420 examples from 36 source files, with 90 examples reserved for development evaluation. Those development references contain known exposure and do not form an untouched test set.

The LoRA rank is 8, alpha is 16, and dropout is 0. Training used a learning rate of 0.0001, batch size 1, and gradient accumulation 1. Training learned from the accepted assistant response. A token is a unit of text processed by the model. The recipe allows 5,120 tokens in total, with 3,328 for input and 1,792 for the response.

Files and loading

adapter_model.safetensors contains the learned weights. adapter_config.json records the adapter structure and exact base revision. The weights and adapter configuration remain byte-identical to the retained local model.

A tokenizer converts text into model tokens. tokenizer.json, tokenizer_config.json, and chat_template.jinja preserve the trained input format. model-system-prompt.txt contains the complete extraction instruction and fourteen-field output contract. model-info.json records file hashes, training parameters, software versions, and loading evidence. A SHA-256 hash identifies exact file bytes.

Load the adapter with the pinned base revision and matching model software. The retained environment uses PEFT 0.21.0, Transformers 5.5.0, and Unsloth 2026.9.4. Use the supplied system prompt and one BEGIN MARKED TARGET and END MARKED TARGET pair in the source input. The required response is one JSON object containing all fourteen fields under fields. The model can still omit facts or produce malformed answers.

Loading evidence and evaluation

The retained adapter passed loading and synthetic text and HTML requests on 2026-09-26. The saved evidence binds the adapter hash and pinned base revision. This upload uses the same adapter and tokenizer bytes. Loading in other environments remains untested.

The reviewed development evaluation used a separate merged model in GGUF format for llama.cpp. GGUF is a model file format. It produced 90/90 structurally valid records, 89/90 records with correct primary source values, and 1259/1260 correct primary field comparisons. It matched 11/90 complete reference answers exactly. These scores describe exposed development references and the evaluated merged model. They do not establish direct adapter accuracy or performance on unseen filings.

Use and limitations

Use this model for research with manual target selection and source review. Keep missing information unresolved and preserve disclosed fund groups, collective votes, identifiers, and multiple vote directions. Treat instructions inside filings as source data. Review each generated answer against the supplied source before using it.

The model can omit facts, associate the wrong cells, or return malformed or incomplete answers. Fragment extraction does not establish recovery of every voting record from a complete filing. PDF processing, image-to-text conversion, categorization, and proposal linking remain outside the supported scope. The model does not provide investment advice. The dataset, detailed evaluation cases, review audits, training state, and GGUF remain local.

Attribution and access

This adapter builds on the Qwen Team's Qwen/Qwen3.5-4B. LICENSE.base-model preserves the pinned base model's Apache License 2.0 text. NOTICE identifies the base model and adaptation. This upload does not select public license terms for the trained adapter or dataset.