Doosra is a cricket analytics app with a local AI copilot: you ask a question, and the model answers it by calling Doosra’s 25 tools (leaderboards, player profiles, match-ups, venues, records, match replays, charts, SQL…) over 11 million balls of Cricsheet data, then writes an answer from the results. This is Qwen3.5 4B fine-tuned on 6,695 of Doosra’s own tool-calling conversations, so the small model does the job reliably, with a prompt about 8x shorter than the app gives general models (about 280 tokens instead of 2,100, plus compact tool schemas), which matters on a laptop where reading the prompt is the slow part. In Doosra’s evaluation it answers 33 of 40 open questions right, against 30 for Qwen3.5 9B and 29 for the base 4B, about 2.5x faster (see Evaluation). v2: rank 32, learning rate 1e-4, 2 epochs, dataset v2; the model Doosra ships. The base model abbreviates the competition and cuts the table to 5 rows nobody asked for. The fine-tuned model makes the exact call; Doosra runs it, and the answer is written from the result (this one is from the test set): > Rashid Khan is top with 49.07 dot % (Bangladesh Premier League, male, death). > — Nasum Ahmed (Khulna Tigers): 47.92…
Модальности:
Генерация текста
Области применения:
Диалог / чат Вызов функций (Tool use) Агенты
Задача: Генерация текста
Автор: Sarthak213
Теги: gguf, qwen3_5, tool-calling, function-calling, agents, cricket, sports-analytics, lora
Лайков: 3 | Загрузок: 399
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.