GX-XinGao/Qwen2.5-7B-R-Select-100k

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 17, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

GX-XinGao/Qwen2.5-7B-R-Select-100k is a 7.6 billion parameter supervised fine-tuned (SFT) model based on Qwen2.5-7B-Base, developed by GX-XinGao. It was trained using the R-Select-100k dataset, which is curated through a robust multi-metric data selection approach. This model excels across general, math, code, and reasoning domains, achieving superior average performance on unseen benchmarks compared to other open-source SFT datasets of similar scale.

Loading preview...

Model Overview

GX-XinGao/Qwen2.5-7B-R-Select-100k is a 7.6 billion parameter supervised fine-tuned (SFT) model built upon the Qwen2.5-7B-Base architecture. Its key differentiator lies in its training data: the R-Select-100k dataset, which comprises 100,000 high-value samples selected using the novel R-Select framework.

R-Select Data Curation

The R-Select framework, detailed in a KDD 2026 paper, addresses the challenge of selecting high-quality SFT data from large, heterogeneous instruction-tuning pools. It moves beyond single-metric filtering by formulating data selection as a multi-metric weight optimization problem. Each sample is annotated with 30 quality metrics (model-based, heuristic, and LLM-as-Judge), which are then clustered and optimized hierarchically using Bayesian optimization with a proxy model (Qwen3-1.7B-Base) and Optuna TPE. This process identifies the top 100K samples from an initial pool of over 3.4 million samples across 21 public datasets, covering diverse domains like reasoning, code, math, and general instruction following.

Performance Highlights

Evaluated with OpenCompass on unseen benchmarks, Qwen2.5-7B-R-Select-100k demonstrates strong performance across general, math, code, and reasoning tasks. It achieves the best reported average performance among compared open-source SFT datasets for both Qwen2.5-7B-Base and Qwen3-8B-Base, despite using a relatively small 100K sample training set. For instance, when applied to Qwen3-8B-Base, it achieves a 69.9 AVG score, outperforming other datasets like MiroMind (69.7) and OmniThought (67.9).

Use Cases

This model is particularly well-suited for applications requiring strong performance across a broad range of tasks, including:

  • General instruction following
  • Mathematical problem-solving
  • Code generation and understanding
  • Complex reasoning tasks

Its efficient data selection process makes it a valuable choice for developers seeking a high-performing model trained on a meticulously curated dataset.

Finetunes
1 models