RiverONE

QCalEval ZeroShot Evaluation Results

Evaluation set: RiverOne_ZeroShot_v1.2_Test (243 samples)

Task	Score	Evaluation Method	Notes
Q1 Chart Understanding	63.63%	LLM Judge (GPT-5)	Programmatic fields (50%) + key_points (50%)
Q2 Chart Classification	86.83%	Exact match	Chart type classification (211/243)
Q3 Key Point Extraction	46.85%	LLM Judge (GPT-5)	Per-item scoring of key_points
Q4 Assessment Judgment	96.30%	Exact match	Assessment classification (234/243)
Q5 Field Extraction	83.55%	Tolerance rules	Field tolerance scoring
Q6 Status Classification	85.60%	Exact match	Status classification (208/243)
Average	77.13%

Quick Start

Requirements

pip install transformers>=4.45 torch>=2.0 aqlm

Loading the Model

import torch
from transformers import AutoTokenizer, AutoConfig
from modeling_riverone_qc import RiverOneQCModel

# Load model config
config = AutoConfig.from_pretrained(
    "path/to/this/model",
    trust_remote_code=True
)

# Load model
model = RiverOneQCModel.from_pretrained(
    "path/to/this/model",
    config=config,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(
    "path/to/this/model",
    trust_remote_code=True
)

Inference Example

from PIL import Image

# Load image
image = Image.open("chart.png").convert("RGB")

# Build conversation
messages = [
    {"role": "user", "content": "<image>\nPlease describe this chart."}
]

# Inference
response = model.chat(
    tokenizer=tokenizer,
    pixel_values=image,
    messages=messages,
    max_new_tokens=512
)
print(response)

Repository Structure

.
├── config.json                    # Model config
├── quant_config.json              # AQLM quantization config
├── generation_config.json         # Generation config
├── model.safetensors              # Model weights (6.9 GB)
├── model.safetensors.index.json   # Weight index
├── tokenizer.json                 # Tokenizer
├── tokenizer_config.json          # Tokenizer config
├── added_tokens.json              # Additional tokens
├── special_tokens_map.json        # Special token mapping
├── chat_template.jinja            # Chat template
├── configuration_riverone_qc.py   # Model configuration class
├── modeling_riverone_qc.py        # Model implementation
├── modeling_ising_vit.py          # ViT implementation
├── processor_config.json          # Processor config
├── preprocessor_config.json       # Preprocessor config
├── video_preprocessor_config.json # Video preprocessor config
└── README.md                      # This file

Citation

This model is based on the following works:

• Qwen3: Qwen Technical Report
• AQLM: Extreme Compression of Large Language Models via Additive Quantization
• InternVL: InternVL: Scaling up Vision Foundation Models
• IsingViT: Quantum-inspired Vision Transformer