prithivMLmods/amazon-LAR-27B-GGUF

> LAR-27B is a 27-billion-parameter open-weights agentic language model developed by Amazon, post-trained from Qwen/Qwen3.8-27B (Apache 2.0) on a license-filtered subset of the SWE-Rebench dataset to solve complex, long-horizon software engineering and terminal-usage tasks. Moving beyond sparse outcome-only supervision, the model is trained with the Lean4AgenticRL framework, which grounds open-ended agentic reinforcement learning in a unified Lean 4 formalization schema. This methodology establishes problem-specific workflows and generates verifiable, multi-level process rewards—capturing low-level tool-call accuracy, mid-level workflow-stage progress, and high-level task outcomes—to guide policy updates via workflow-stage-normalized Group Relative Policy Optimization (GRPO). Through this formal, multi-granular reward structure, LAR-27B sets a strong benchmark for its parameter class, achieving 84.2% on SWE-Bench-Verified, 68.2% on TerminalBench-2.0, and 77.3% on TerminalBench-2.1. LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp

Модальности:
Генерация текста

Области применения:
Диалог / чат

Задача: Генерация текста
Автор: prithivMLmods
Теги: gguf, text-generation-inference, llama-cpp, amazon, agent, Reinforcement Learning, en, endpoints_compatible
Лайков: 3  |  Загрузок: 591

Открыть на HuggingFace →

Описание основано на материалах HuggingFace. Перевод выполнен автоматически.