87 lines
3.2 KiB
Markdown
87 lines
3.2 KiB
Markdown
|
---
|
||
|
|
library_name: transformers
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
base_model: Qwen/Qwen3-8B
|
||
|
|
---
|
||
|
|
|
||
|
|
# OpenRubrics/RubricARROW-8B-Rubric
|
||
|
|
|
||
|
|
This is an 8B rubric generation model, introduced in the paper [RUBRIC-ARROW: Alternating Pointwise Rubric Reward Modeling for LLM Post-training in Non-verifiable Domains](https://huggingface.co/papers/2605.29156).
|
||
|
|
|
||
|
|
It is finetuned from [Qwen3/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B).
|
||
|
|
|
||
|
|
# Usage
|
||
|
|
|
||
|
|
```python
|
||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
|
model_id = "OpenRubrics/RubricARROW-8B-Rubric"
|
||
|
|
tok = AutoTokenizer.from_pretrained(model_id, use_fast=True)
|
||
|
|
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto")
|
||
|
|
```
|
||
|
|
|
||
|
|
To evaluate the model, please use the following format to build up message.
|
||
|
|
|
||
|
|
```python
|
||
|
|
RUBRIC_PROMPT_TEMPLATE = (
|
||
|
|
"Your task is to extract a set of rubric-style instructions from a user's request.\n"
|
||
|
|
"These rubrics will be used as evaluation criteria to check if a response fully meets the request.\n"
|
||
|
|
"Every rubric item must be a universal principle. If any rubric still contains topic-specific references (e.g., names, places, myths, numbers, historical facts), it is automatically invalid.\n"
|
||
|
|
"\n"
|
||
|
|
"- **Two Distinct Categories:**\n"
|
||
|
|
" - [Hard Rule]: Derived strictly from explicit requirements stated in the <request> (format, length, structure, forbidden/required elements, etc.).\n"
|
||
|
|
" - [Principle]: Derived by abstracting any concrete cues into domain-agnostic quality criteria (e.g., clarity, correctness, sound reasoning, pedagogy).\n"
|
||
|
|
"\n"
|
||
|
|
"- **Comprehensiveness:**\n"
|
||
|
|
" The rubric must cover all critical aspects implied by the request and examples, including explicit requirements and implicit quality standards.\n"
|
||
|
|
"\n"
|
||
|
|
"- **Conciseness & Uniqueness:**\n"
|
||
|
|
" Each rubric must capture a distinct evaluation criterion. Overlapping or redundant criteria must be merged into a single rubric. Wording must be precise and free of repetition.\n"
|
||
|
|
"\n"
|
||
|
|
"- **Format Requirements:**\n"
|
||
|
|
" - Use a numbered list.\n"
|
||
|
|
" - Each item starts with \"The response\" phrased in third person.\n"
|
||
|
|
" - Append [Hard Rule] or [Principle] at the end of each item.\n"
|
||
|
|
" - Do not include reasoning, explanations, or examples in the final output—only the rubrics.\n"
|
||
|
|
"\n"
|
||
|
|
"Here is the request:\n"
|
||
|
|
"{prompt}\n"
|
||
|
|
"\n"
|
||
|
|
"Please generate the rubrics for the above request."
|
||
|
|
)
|
||
|
|
|
||
|
|
user_text = RUBRIC_PROMPT_TEMPLATE.format(
|
||
|
|
prompt=instruction,
|
||
|
|
)
|
||
|
|
|
||
|
|
messages_list = [
|
||
|
|
{"role": "user", "content": user_text},
|
||
|
|
]
|
||
|
|
|
||
|
|
message = tok.apply_chat_template(
|
||
|
|
messages_list,
|
||
|
|
tokenize=False,
|
||
|
|
add_generation_prompt=True,
|
||
|
|
enable_thinking=False
|
||
|
|
)
|
||
|
|
|
||
|
|
# Remaining step: Use either HF or vLLM for evaluation.
|
||
|
|
# ...
|
||
|
|
# ...
|
||
|
|
```
|
||
|
|
|
||
|
|
|
||
|
|
# Citation
|
||
|
|
|
||
|
|
If you find our work helpful, please consider citing our paper:
|
||
|
|
|
||
|
|
```bibtex
|
||
|
|
@misc{jiang2026rubric,
|
||
|
|
title={RUBRIC-ARROW: Alternating Pointwise Rubric Reward Modeling for LLM Post-training in Non-verifiable Domains},
|
||
|
|
author={Haoxiang Jiang and Zihan Dong and Tianci Liu and Wanying Wang and Ran Xu and Tony Yu and Linjun Zhang and Haoyu Wang},
|
||
|
|
year={2026},
|
||
|
|
eprint={2605.29156},
|
||
|
|
archivePrefix={arXiv},
|
||
|
|
primaryClass={cs.LG},
|
||
|
|
url={https://arxiv.org/abs/2605.29156},
|
||
|
|
}
|
||
|
|
```
|