A PEFT LoRA adapter fine-tuned on top of google/gemma-4-e4b-it using the Crownelius/Opus-4.6-Reasoning-2100x-formatted dataset. — structured step-by-step reasoning — logic puzzles — planning and decomposition — algorithm explanations — conceptual problem solving — code reasoning workflows The strongest improvements are visible on: — multi-step logic puzzles — algorithm design explanations — state-tracking tasks — proof-style conceptual reasoning The adapter shows strongest gains on deliberate decomposition, planning, and educational reasoning prompts. — training loss: 192.38 — validation loss: 11.95 — entropy: 3.91 — mean token accuracy: 0.0462 — train runtime: 5783 seconds — train rows: 2010 — validation rows: 106 — structured reasoning — decomposition — planning — conceptual explanation For exact equation solving, outputs may sometimes over-interpret terse symbolic prompts. This adapter is a derivative of Gemma 4 and follows the Gemma license terms.
Модальности:
Генерация текста
Области применения:
Генерация кода Логика и рассуждение Диалог / чат
Задача: Генерация текста
Автор: krishnamraja13
Теги: gemma, peft, lora, reasoning, puzzle-solving, code, conversational, endpoints_compatible
Лайков: 3 | Загрузок: 0
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
