RC-OPD

RC-OPD provides Qwen3-based mathematical reasoning models in 1.7B, 4B, and 8B sizes, trained with targeted feedback on reasoning errors. The released weights are LoRA adapters for the corresponding Qwen3 base models.

Evaluation results

All scores below are Avg@4 accuracy (%) on the full AIME24, AIME25, and HMMT25 benchmark sets used by the OPSD evaluator. Each benchmark contains 30 problems, with four generated solutions per problem (120 solutions per benchmark). The results were checked against all nine saved evaluation outputs and their per-generation correctness labels; evaluation was not rerun for this release.

Size	AIME24	AIME25	HMMT25	Mean	Adapter
1.7B	54.17	47.50	30.83	44.17	RC-OPD-1.7B
4B	78.33	69.17	50.83	66.11	RC-OPD-4B
8B	78.33	76.67	45.83	66.94	RC-OPD-8B

Metric: Avg@4 (%) = 100 × total correct generated solutions / (30 × 4). The Mean column is the arithmetic mean of the three benchmark scores, computed before rounding. Pass@4 (at least one correct solution per problem) and majority-vote accuracy are separate metrics and are not the scores in this table. The 8B adapter has the highest three-benchmark mean among these three released variants.

Evaluation prompt

Each problem is submitted as a single user message with the exact following content:

{problem}

Please reason step by step, and put your final answer within \boxed{}.

{problem} is replaced by the benchmark problem text. There is no additional system message or few-shot example. The user message is formatted using the matching Qwen3 tokenizer's chat template:

messages = [{
    "role": "user",
    "content": problem + "\n\nPlease reason step by step, and put your final answer within \\boxed{}.",
}]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True,
)

Decoding and scoring

Parameter	Value
Inference engine	vLLM with the selected LoRA adapter
Thinking mode	Enabled
Samples per problem (n / val_n)	4
Temperature	1.0
top_p	0.95
top_k	-1 (disabled)
min_p	0.0
Presence penalty	0.0
Maximum generated tokens	38,912
Maximum context length	40,960

Dataset identifiers in the evaluation implementation are HuggingFaceH4/aime_2024, yentinglin/aime_2025, and MathArena/hmmt_feb_2025. Each contributes its full 30-problem set; the loader uses the dataset's train split label for these benchmark questions.

Scoring follows the original OPSD evaluator: extract the final \boxed{...} answer and check it with math_verify. If the verifier raises an exception, the evaluator falls back to normalized string comparison. A missing boxed answer is scored incorrect. These reported results use this evaluator without additional LLM regrading.

Files and integrity

Each adapter directory contains adapter_model.safetensors, adapter_config.json, and a short model card. Adapter weights retain their original SHA256 hashes. The configs reference the matching public Qwen3 base-model identifiers for portability. Base-model weights and tokenizers must be loaded separately.

checkpoint_manifest.json records the released file sizes and SHA256 hashes.

Loading

Load the matching base model and its adapter as below. Change both base_id and subfolder to select another size.

from huggingface_hub import snapshot_download
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen3-1.7B"
subfolder = "RC-OPD-1.7B"

checkpoint_root = snapshot_download(
    repo_id="starrylay/RC-OPD",
    allow_patterns=[f"{subfolder}/adapter_config.json", f"{subfolder}/adapter_model.safetensors"],
)
tokenizer = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(
    base_id, torch_dtype="auto", device_map="auto",
)
model = PeftModel.from_pretrained(base, f"{checkpoint_root}/{subfolder}")
model.eval()

Directory	Base model
RC-OPD-1.7B	Qwen/Qwen3-1.7B
RC-OPD-4B	Qwen/Qwen3-4B
RC-OPD-8B	Qwen/Qwen3-8B