AccelerateScience/Qwen3-14B-gwb-press-conference-sft-merged

Model Overview

AccelerateScience/Qwen3-14B-gwb-press-conference-sft-merged is a 14 billion parameter language model built upon the Qwen3 architecture. This model represents a merged version of an adapter-only fine-tuned variant, specifically developed by AccelerateScience. Its designation, 'gwb-press-conference-sft', indicates a specialized fine-tuning process likely targeting tasks related to press conference content, such as summarization, generation, or analysis within that domain.

Key Characteristics

• Parameter Count: 14 billion parameters, offering substantial capacity for complex language understanding and generation.
• Context Length: Supports a context window of 32768 tokens, enabling processing of lengthy inputs relevant to detailed documents or conversations.
• Specialized Fine-tuning: The model is a supervised fine-tuned (SFT) version, suggesting optimization for specific instruction-following or text generation tasks within its target domain.
• Validation Score: Achieves a validation score of 0.120 [0.111, 0.130], providing an indication of its performance on the validation set used during its development.

Generation Configuration

The model's recommended generation settings include:

• do_sample: True
• temperature: 0.7
• top_p: 1.0
• top_k: 0
• repetition_penalty: 1.0
• max_new_tokens: 1024

These settings are designed to balance creativity and coherence in generated outputs.

Good For

• Domain-Specific Text Generation: Ideal for applications requiring text generation or analysis within the context of press conferences or similar formal communication events.
• Specialized SFT Tasks: Suitable for tasks where a supervised fine-tuned model with a large context window is beneficial.