Stage-org/4b-300-hard-27b-z-iter2-epoch3

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 22, 2026Architecture:Transformer Featherless Exclusive Cold

Stage-org/4b-300-hard-27b-z-iter2-epoch3 is a 4.5 billion parameter language model developed by Stage-org, fine-tuned using a reinforcement learning (RL) method. This model is trained on the Stage-org/4b-300-hard-27b-z-iter2 dataset with a context length of 32768 tokens. It is optimized for complex tasks, leveraging an open-ended judge for evaluation and advanced generation parameters, making it suitable for applications requiring sophisticated reasoning and problem-solving capabilities.

Loading preview...

Model Overview

Stage-org/4b-300-hard-27b-z-iter2-epoch3 is a 4.5 billion parameter language model developed by Stage-org. This model has undergone reinforcement learning (RL) training, specifically using the Stage-org/4b-300-hard-27b-z-iter2 dataset. It features a substantial context length of 32768 tokens, enabling it to process and generate extensive sequences of text.

Key Capabilities

  • Reinforcement Learning (RL) Optimization: The model is fine-tuned using an RL method, indicating a focus on improving performance through iterative feedback and reward mechanisms.
  • Advanced Inference Configuration: Utilizes flash_attention_2 for efficient attention mechanisms and vllm_extra with language_model_only mode, suggesting optimized inference for language tasks.
  • Sophisticated Judging Mechanism: Incorporates an open_ended_judge with gpt-5.6-luna as the evaluation model, featuring retry logic and reasoning effort settings, which implies a focus on high-quality, nuanced outputs.
  • Configurable Generation: Supports generation with parameters like temperature (0.9), max tokens (4096), and top_p (1.0), along with an enable_thinking option, indicating capabilities for complex and creative text generation.

Good For

  • Complex Problem Solving: The integration of an open-ended judge and reasoning effort in generation suggests suitability for tasks requiring intricate problem-solving and nuanced responses.
  • Research and Development in RL: Given its explicit RL training provenance and detailed configuration, this model is well-suited for researchers and developers exploring advanced RL applications in language models.
  • Applications Requiring High-Quality Output: The use of a powerful external judge for evaluation points to an emphasis on generating high-quality, contextually appropriate, and well-reasoned outputs.