Agent-style chat model obtained by LoRA fine-tuning openbmb/MiniCPM5-2B (2.57B parameters, Llama architecture, 130,560-token vocabulary) on agent traces from saidutta69/fable-5-premium-v2 (openaichat` config), distilled from Claude Fable-5 and other frontier models. LoRA adapters (50.2M parameters, 1.96%) are merged into fp16 weights below, so the model loads and serves like any standard causal LM — no PEFT code required. The chat template is Qwen3-style (/, with an empty block on assistant turns) and ships with the tokenizer. Intended for research and prototyping of tool-using conversational agents: multi-step instruction following, function/tool calling, and agentic dialogue in English and Chinese. Out of scope: high-stakes decisions (medical, legal, financial), unsupervised autonomous actions, and any use as a source of factual truth — like all models distilled from frontier traces, it can hallucinate tools, arguments, and facts. Training saw truncated contexts (see below), so very long agentic episodes exceed what was trained. Safety alignment is inherited from the base model and the distilled data only; no dedicated red-teaming was performed. — Source:…
Модальности:
Генерация текста
Области применения:
Диалог / чат Вызов функций (Tool use)
Задача: Генерация текста
Автор: SauravMahalik
Теги: llama, minicpm, sft, lora, fable-5, tool-use, agent-traces, chat
Лайков: 3 | Загрузок: 165
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
