SynGen-14B: Synthetic Reasoning Generation

SynGen-14B is a 14 billion parameter language model developed by qingy2024, built upon the Qwen/Qwen3-14B architecture. Its core purpose is to generate synthetic grounded reasoning, acting as an intermediary step between a user's prompt and the model's final output. This capability is particularly useful for dataset modification, allowing users to convert standard chat datasets into reasoning-rich datasets, emulating the style of models like DeepSeek R1 or OpenAI's GPT OSS.

Key Capabilities

• Synthetic Reasoning Generation: Inserts explicit reasoning steps into model outputs.
• Dataset Transformation: Converts existing chat datasets into reasoning-focused datasets.
• Style Emulation: Can generate reasoning in the style of DeepSeek R1 or GPT OSS.
• Flexible Prompt Format: Utilizes specific tags (<reasoning_style>, <system_prompt>, <user>, <assistant>, <think>) for structured input and output.

Training Details

• Base Model: Qwen/Qwen3-14B
• Training Method: Full Fine-Tune (FFT) over 2 epochs.
• Dataset: Based on Pinkstack/syngen-reasoning-0.6b-dataset.
• Total Tokens Trained: Approximately 211 million tokens.

Recommended Usage

For optimal performance and to prevent repetitive loops, it is recommended to use a temperature = 1.0 with default sampler settings when interacting with SynGen-14B.