Skip to content

README.md

Latest commit

 

History

History
94 lines (78 loc) · 4.42 KB

README.md

File metadata and controls

94 lines (78 loc) · 4.42 KB
Raw

RCPS/RADP parser-training adapters

EMNLP 2026 Release

This release contains the nine final LoRA adapters evaluated in Retrieval-Conditional Parsing Score (RCPS): Choosing Document Parsers by Retrieval, Not by Appearance, published in the Proceedings of the EMNLP 2026 Industry Track (ACL Anthology). It includes the RADP-aux lambda sweep, RADP-DPO R1--R3, RADP-Distill, and the SimPO control. These checkpoints support the paper's secondary negative-result study; they are not a new production recommendation. The paper, evaluation code, and deterministic release builder live in wigtn/WigtnOCR-RADP.

The main evaluation base is Wigtn/Qwen3-VL-2B-WigtnOCR. RADP-DPO, RADP-Distill, and SimPO were trained from that base. RADP-aux was trained from Qwen/Qwen3-VL-2B-Instruct and then evaluated in the paper by applying its adapter to the Prod base. The per-variant adapter_config.json records the training base, while manifest.json records both training and evaluation bases.

Where these adapters fit

Step Repository What it holds
1 wigtn/wigtnOCR-v1 Distillation of Qwen3-VL-30B pseudo-labels into Wigtn/Qwen3-VL-2B-WigtnOCR, the base these adapters load onto (called Prod in the paper)
2 wigtn/WigtnOCR-RADP The RCPS paper, evaluation code, the 663-Q–A KoGovDoc-RAG probe, and result artifacts
3 This repository The nine LoRA adapters from the paper's secondary parser-training study

Variants

Paper label Subfolder
RADP-aux lambda=0.0 radp-aux/lambda-0.0
RADP-aux lambda=0.1 radp-aux/lambda-0.1
RADP-aux lambda=0.3 radp-aux/lambda-0.3
RADP-aux lambda=0.5 radp-aux/lambda-0.5
RADP-DPO-R1 radp-dpo/r1
RADP-DPO-R2 radp-dpo/r2
RADP-DPO-R3 radp-dpo/r3
RADP-Distill radp-distill
RADP-SimPO radp-simpo

Loading an evaluated adapter

Download and extract the audited release first:

curl -L -o RCPS-RADP-Adapters-v1.tar.gz "https://github.com/wigtn/RCPS-RADP-Adapters/releases/download/v1.0.0/RCPS-RADP-Adapters-v1.tar.gz"
mkdir -p /path/to/RCPS-RADP-Adapters
tar -xzf RCPS-RADP-Adapters-v1.tar.gz -C /path/to/RCPS-RADP-Adapters

The release tarball is 358,728,348 bytes with SHA-256 22b4a0d4f5560f4d7c31633a1c885bbe3568a2bef841bfd94bf2407c339f08c6.

from peft import PeftModel
from transformers import Qwen3VLForConditionalGeneration

base = Qwen3VLForConditionalGeneration.from_pretrained(
    "Wigtn/Qwen3-VL-2B-WigtnOCR", torch_dtype="auto", device_map="auto"
)
model = PeftModel.from_pretrained(
    base,
    "/path/to/RCPS-RADP-Adapters/radp-dpo/r2",
)

manifest.json gives original and release hashes, executed hyperparameters, source-log hashes, and the source repository paths for every variant. Structured trainer_state.json files are included where the Transformers trainer produced them. Optimizer states, training images, and preference-pair text are excluded. The repository release builder verifies the original weights before staging. See the main repository's CHECKPOINT_RELEASE.md for the inclusion policy and CPU-only verification command.

Citation

@inproceedings{son2026rcps,
  title     = {Retrieval-Conditional Parsing Score ({RCPS}): Choosing Document Parsers by Retrieval, Not by Appearance},
  author    = {Son, Sang-Woo and Kim, Hyeong-seob and Kim, Hyeonsang and Cho, Hyun-woo and Kim, Jinmo},
  editor    = {Atanasova, Pepa and Haffari, Gholamreza and Wang, Minghan},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track},
  month     = oct,
  year      = {2026},
  address   = {Budapest, Hungary},
  publisher = {Association for Computational Linguistics},
  pages     = {1531--1544},
  url       = {https://aclanthology.org/2026.emnlp-industry.90/}
}