metadata
license: apache-2.0
language:
- en
library_name: transformers
tags:
- math
- cycleqd
- qwen2.5
- merge
datasets:
- HLE-Math
base_model:
- Qwen/Qwen2.5-7B
- Qwen/Qwen2.5-7B-Instruct
model-index:
- name: CycleQD-Qwen2.5-7B-Math-Alpha020
results:
- task:
type: mathematical-reasoning
name: HLE Math
dataset:
name: HLE Math
type: hle-math
metrics:
- type: accuracy
value: 14.29
name: accuracy
verified: true
details: 5 correct out of 35 questions
CycleQD Qwen2.5-7B Math α=0.20
This model is a CycleQD-merged version of Qwen2.5-7B, specifically tuned for mathematical reasoning tasks.
Model Details
- Base Models: Qwen2.5-7B and Qwen2.5-7B-Instruct
- Merge Method: Linear interpolation with α=0.20
- Formula:
(1-0.20) × Qwen2.5-7B + 0.20 × Qwen2.5-7B-Instruct - Created: August 14, 2025
- Model Size: 7B parameters (~15GB)
Performance
This model achieved significant improvement on the HLE Math evaluation:
- Accuracy: 14.29% (5/35 questions correct)
- Improvement: 5x improvement from baseline (1/35 → 5/35)
- Evaluation: HLE Math category
- Judge Model: Qwen2.5-32B-Instruct
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("You3dimgeo/cycleqd-qwen25-7b-math-alpha020")
tokenizer = AutoTokenizer.from_pretrained("You3dimgeo/cycleqd-qwen25-7b-math-alpha020")
# Use for mathematical reasoning
prompt = "Solve the equation: 2x^2 + 5x - 3 = 0"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=200)
print(tokenizer.decode(outputs[0]))
Training/Merge Details
This model was created using the CycleQD (Cycle Quality-Diversity) approach:
- Started with Qwen2.5-7B base model
- Applied CycleQD optimization for mathematical reasoning
- Merged with instruction-tuned variant using α=0.20
- Evaluated on HLE Math benchmark
License
This model is licensed under Apache 2.0, following the original Qwen2.5 license.