Search Results for "phase"

...The innovation lies in its training methodology: the team uses what they call the Spectrum-to-Signal Principle (SSP), where a first stage emphasizes diversity of reasoning paths (the “spectrum” phase) and a second stage uses reinforcement techniques (the “signal” phase) to refine toward correctness and strong reasoning. The result is a model that outpaces many much larger models on domain-specific benchmarks, demonstrating that smaller models, if trained carefully and with the right objectives, can achieve high performance in reasoning-centric tasks.

...The project explores the idea that instead of stacking hundreds of unique transformer layers, a smaller set of layers can be reused iteratively during inference to achieve deeper reasoning without increasing parameter count. It divides computation into three main stages, including a pre-processing phase, a looped recurrent reasoning block, and a final output refinement stage, creating a structured pipeline for inference. The architecture incorporates advanced techniques such as mixture-of-experts routing, adaptive computation time, and multiple attention mechanisms to dynamically allocate compute where needed. It is highly configurable through a centralized configuration system, allowing experimentation with different architectural parameters such as loop depth, attention type.

This project is developed based on the commercially available large model Llama-2 released by Meta. It is the second phase of the Chinese LLaMA&Alpaca large model project. The Chinese LLaMA-2 base model and the Alpaca-2 instruction fine-tuning large model are open-sourced. These models expand and optimize the Chinese vocabulary on the basis of the original Llama-2, use large-scale Chinese data for incremental pre-training, and further improve the basic semantics and command understanding of Chinese. ...

...Data are stored as newline-delimited JSONL files tracked with Git LFS, where each line is a full solution sample that can contain many step-level labels and rich metadata such as labeler UUIDs, timestamps, generation identifiers, and quality-control flags. Each labeled step can include multiple candidate completions with ratings of -1, 0, or +1, optional human-written corrections (phase 1), and a chosen completion index, along with a final finish reason such as found_error, solution, bad_problem, or give_up.

...Built on Qwen 2.5-VL-7B and fine-tuned with synthetic Doc → Reasoning → Markdown examples, it generates thinking tokens before producing the final Markdown to better handle complex layouts and tables. It uses a two-phase training process: supervised fine-tuning (SFT) followed by reinforcement learning (GRPO) with a layout-centric reward for accuracy on challenging documents. The model excels at non-standard layouts and complex table structures, outperforming non-reasoning OCR systems like GPT-4o and OCRFlux, and competing with large closed-source reasoning models like Gemini 2.5. ...

Jan-v1-edge is a lightweight agentic language model developed by JanHQ, designed for fast and reliable on-device execution. It is the second release in the Jan Family and was distilled from the larger Jan-v1 model, retaining strong reasoning and problem-solving capabilities while reducing its computational footprint. The model was refined through a two-stage post-training process: Supervised Fine-Tuning (SFT) to transfer knowledge from Jan-v1, followed by Reinforcement Learning with...