open-thoughts/OpenThoughts-Agent-SFT-316
收藏资源简介:
OpenThoughts-Agent-SFT-316是一个开源数据集,属于OpenThoughts-Agent项目的一部分,旨在策划用于训练智能体的高质量数据集。该数据集包含316个示例,是SFT(监督微调)扩展阶梯中的一个点(其他规模包括1K、3.16K等)。数据由(任务,智能体轨迹)对组成,用于微调OpenThinkerAgent-8B-SFT-316和OpenThinkerAgent-32B-SFT-316模型。任务来源于四个主要任务源:SWE-Smith、StackExchange-SuperUser、StackExchange-Tezos(通过合成增强以扩展任务多样性)和IssueTasks。智能体轨迹由GLM-4.7-AWQ模型作为教师在terminus-2框架内生成,并过滤了至少包含5个模型轮次的轨迹。数据集字段包括conversations(多轮智能体轨迹消息)、task(任务描述)、trace_source(任务来源)、agent/model/model_provider(元数据)以及result/episode/run_id/trial_name/date(记录信息)。行数为316,教师模型为GLM-4.7-AWQ,框架为terminus-2。
OpenThoughts-Agent-SFT-316 is an open-source dataset curated as part of the OpenThoughts-Agent project to provide high-quality data for training agents. It contains 316 examples, representing one point in the SFT scaling ladder (with other sizes including 1K, 3.16K, etc.). The dataset consists of (task, agent-trajectory) pairs used to fine-tune models such as OpenThinkerAgent-8B-SFT-316 and OpenThinkerAgent-32B-SFT-316. Tasks are drawn from four primary sources: SWE-Smith, StackExchange-SuperUser, StackExchange-Tezos (synthetically augmented to expand task diversity), and IssueTasks. Agentic trajectories are generated by the GLM-4.7-AWQ model acting as the teacher within the terminus-2 harness, filtered to include traces with at least 5 model turns. Fields include conversations (multi-turn agent trajectory messages), task (task description), trace_source (originating task source), agent/model/model_provider (metadata), and result/episode/run_id/trial_name/date (bookkeeping information). It has 316 rows, with the teacher model being GLM-4.7-AWQ and the harness being terminus-2.




