nemotron-code-oracle-qwen3.5-122b-opencode-sft
收藏资源简介:
nemotron-code-oracle数据集是一个专为监督微调(SFT)设计的多轮智能体对话数据集。其内容来源于使用Qwen/Qwen3.5-122B-A10B-FP8模型(支持131k上下文,并开启验证器)生成的opencode编码智能体轨迹。数据集的核心价值在于其忠实还原了模型在推理和工具调用过程中的原始输出:每个助手(assistant)的回复轮次均直接从服务引擎的字面补全token IDs解码而来,完整保留了模型内部的真实推理过程(以`<think>…</think>`标记)以及其原生的opencode工具调用格式(`<tool_call><function=…><parameter=…>`)。系统提示和任务轮次从第一步的提示解码,而工具调用和观察结果轮次则来自原始轨迹。数据集中不包含字面token的行已被过滤。数据集主要包含以下字段:`conversations`字段采用ShareGPT格式(即`[{from: system|human|gpt, value}]`的列表);`text`字段为扁平化的完整对话字符串,使用Qwen的`<|im_start|>`标记,并关键性地在每一轮都保留了推理内容(这与标准聊天模板会从历史中剥离`<think>`的做法不同);此外还有`task`(任务描述)和`num_turns`(对话轮数)字段。该数据集适用于训练或微调能够进行复杂、多轮、工具增强型代码相关任务的对话模型。
The nemotron-code-oracle dataset is a multi-turn agent dialogue dataset specifically designed for supervised fine-tuning (SFT). Its content is derived from opencode-encoded agent trajectories generated using the Qwen/Qwen3.5-122B-A10B-FP8 model (which supports 131k context length and has the validator enabled). The core value of the dataset lies in its faithful reproduction of the models original output during reasoning and tool invocation: each assistant response turn is directly decoded from the literal completion token IDs of the serving engine, fully preserving the models internal reasoning process (marked with `<think>…</think>`) and its native opencode tool call format (`<tool_call><function=…><parameter=…>`). System prompts and task turns are decoded from the first-step prompt, while tool call and observation turns come from the original trajectory. Rows containing literal tokens have been filtered out. The dataset primarily includes the following fields: the `conversations` field uses the ShareGPT format (i.e., a list of `[{from: system|human|gpt, value}]`); the `text` field is a flattened complete dialogue string using Qwens `<|im_start|>` tokens, and crucially retains reasoning content in every turn (unlike standard chat templates that strip `<think>` from history); additionally, there are `task` (task description) and `num_turns` (number of dialogue turns) fields. This dataset is suitable for training or fine-tuning dialogue models capable of handling complex, multi-turn, tool-enhanced code-related tasks.
nemotron-code-oracle · Qwen3.5-122B-A10B · opencode · SFT 数据集详情
数据集概述
该数据集包含从 opencode 编码智能体轨迹中蒸馏得到的多轮智能体 SFT(监督微调)对话,对话由 Qwen/Qwen3.5-122B-A10B-FP8 模型(131k 上下文,验证器开启)生成。
数据生成方式
- 每个 assistant 轮次直接从推理引擎的完整完成 token ID 中逐字解码(RecordProxy 捕获),使用与
Qwen/Qwen3.5-122B-A10B-FP8完全相同的分词器。 - 因此数据保留了模型真实的推理过程(
<think>…</think>)和原生 opencode 工具调用格式(<tool_call><function=…><parameter=…>)。 - 系统轮次和任务轮次从第一步提示中解码;工具/观察轮次从轨迹中获取。
- 没有字面量 token 的行会被丢弃。
数据字段
conversations:ShareGPT 格式[{from: system|human|gpt, value}]text:使用 Qwen<|im_start|>标记的扁平化对话。每个轮次均保留推理过程(标准聊天模板会从历史中移除<think>,但该字符串不会,因此每个轮次的推理过程可用于 SFT)task:任务描述num_turns:对话轮次数
来源与转换
- 源轨迹:
penfever/nemotron-code-oracle-filtered-qwen3.5-122b-131k-opencode-traces - 转换方式:将字面量完成 token 通过
AutoTokenizer.from_pretrained("Qwen/Qwen3.5-122B-A10B-FP8")转换为文本
标签
sft、agent-traces、code、opencode- 语言:英语(en)




