laya-scene-graph-papers

A Laya checkpoint fine-tuned to classify arXiv papers that mention "scene graph" into the two-level category tree of all-scene-graph-papers.

Usage

Input is Title: ...\nAbstract: .... questions.json holds one category question and one sub:<category> question per category; take the subcategory answer of the predicted category.

import json, laya
from huggingface_hub import hf_hub_download

repo = "Blu-y/laya-scene-graph-papers"
agent = laya.load(repo)
questions = json.load(open(hf_hub_download(repo, "questions.json")))
r = agent.predict(f"Title: {title}\nAbstract: {abstract}", questions, max_len=768, head_max_len=256)
cat = r["answers"]["category"]["choice"]
sub = r["answers"]["sub:" + cat]["choice"]

Training

• Base: convaiinnovations/laya (English, ModernBERT-large)
• Data: 511 human-labeled papers + 411 papers labeled by Claude with high confidence
• 6 epochs, soft cross-entropy, option-order shuffling, max_len 768 / head_max_len 256

Evaluation

Held-out 154 human-labeled papers, with a model trained on the remaining 357 human labels + the same 411 Claude labels (this released checkpoint also trains on those 154):

	category	category + subcategory
laya zero-shot	62.3%	40.3%
this recipe	77.3%	66.2%
Claude (reference)	81.2%	66.2%