genrec-dataset
收藏资源简介:
genrec-dataset 是一个用于 SIDReasoner 的第二阶段推理数据集,包含由 Azure OpenAI GPT-5.4 生成的链式思维(chain-of-thought)推理轨迹。该数据集专为视频游戏领域的推理任务设计,旨在替代上游的推理轨迹,提供更高质量、多样化的推理过程数据。数据集包含多个子集(config),每个子集对应一个独立的 Parquet 文件,且仅包含训练集(train)。子集分为不同的 prompt 版本:v4、v5 和 v6,每个版本使用不同的生成提示模板。其中,共享相同 prompt 版本的子集(如 Video_Games_reasoning、Video_Games_reasoning_sample2、Video_Games_reasoning_sample3)是从相同的 49,133 行原始数据中以温度 1 独立采样得到的,行通过 _source_index 字段对齐,且推理文本不重复,适合用于多数投票、难度分层或数据增强。不同版本子集的推理文本格式有所差异:v4 生成 4-8 句的分析师独白(证据→兴趣→机制→预测);v5 生成 1-6 句的基于证据的推理,禁止复述完整历史;v6 生成 4-8 句的分析师独白(120-180 词),并额外包含物品描述,以便通过具体属性表征语义 ID 而不提及标题。所有 prompt 输出的每条记录均为严格的 JSON 对象,包含三个字段:predictability(可预测性)、transition_type(转移类型)和 reasoning(推理过程)。该数据集适用于训练和评估推荐系统中的推理模型,特别是基于思维链的强化学习(如 GRPO)训练。此外,数据集仓库还提供了训练好的模型检查点,包括 Stage-2 CoT 和 Stage-3 RL 检查点,可直接加载使用。
The genrec-dataset is a second-stage reasoning dataset for SIDReasoner, containing chain-of-thought reasoning trajectories generated by Azure OpenAI GPT-5.4. It is specifically designed for reasoning tasks in the video game domain, aiming to replace upstream reasoning trajectories with higher-quality and more diverse reasoning process data. The dataset includes multiple subsets (configs), each corresponding to an independent Parquet file, and only contains a training set (train). The subsets are divided into different prompt versions: v4, v5, and v6, each using a different generation prompt template. Subsets sharing the same prompt version (e.g., Video_Games_reasoning, Video_Games_reasoning_sample2, Video_Games_reasoning_sample3) are independently sampled from the same 49,133 rows of raw data at temperature 1, with rows aligned via the _source_index field, and the reasoning texts are non-repetitive, making them suitable for majority voting, difficulty stratification, or data augmentation. The reasoning text formats differ across versions: v4 generates 4-8 sentences of analyst monologue (evidence → interest → mechanism → prediction); v5 generates 1-6 sentences of evidence-based reasoning, prohibiting restating the full history; v6 generates 4-8 sentences of analyst monologue (120-180 words) with additional item descriptions, so that semantic IDs are characterized by specific attributes without mentioning titles. Each record output by all prompts is a strict JSON object containing three fields: predictability, transition_type, and reasoning. The dataset is suitable for training and evaluating reasoning models in recommendation systems, especially for chain-of-thought-based reinforcement learning (e.g., GRPO) training. Additionally, the dataset repository provides trained model checkpoints, including Stage-2 CoT and Stage-3 RL checkpoints, which can be loaded directly for use.
genrec-dataset 数据集概述
基本信息
- 数据集名称:genrec-dataset
- 所属仓库:
mengdanzhu/genrec-dataset - 用途:为 SIDReasoner 提供的第二阶段推理数据,其中
reasoning_path列包含使用 Azure OpenAI GPT-5.4 生成的思维链轨迹,替代了上游轨迹。
数据集结构
配置与数据划分
数据集包含 10 个子配置(config),每个配置仅包含 train 分割,数据文件以 Parquet 格式存储:
| 配置名 | 分割 |
|---|---|
Video_Games_reasoning |
train |
Video_Games_reasoning_sample2 |
train |
Video_Games_reasoning_sample3 |
train |
Video_Games_reasoning_v4 |
train |
Video_Games_reasoning_v5 |
train |
Video_Games_reasoning_v6 |
train |
Video_Games_reasoning_v6_sample2 |
train |
Video_Games_reasoning_v6_sample3 |
train |
Video_Games_reasoning_v6_sample4 |
train |
Video_Games_reasoning_v6_sample5 |
train |
子集特性
- 共享相同提示词版本的子集(如
Video_Games_reasoning及其 sample2、sample3;Video_Games_reasoning_v6及其 sample2 至 sample5)均为对同一 49,133 行数据在不同 temperature=1 条件下的独立采样。 - 子集之间不重复推理文本,可通过
_source_index列进行行级对齐。 - 适用于多数投票、难度分层或数据增强等场景。
提示词版本说明
子集的生成模板位于 prompts/ 目录下,每个 prompt_v<N>.txt 对应生成 *_reasoning_v<N> 子集:
| 提示词文件 | 对应的子集 |
|---|---|
prompts/prompt_v4.txt |
Video_Games_reasoning_v4 |
prompts/prompt_v5.txt |
Video_Games_reasoning_v5 |
prompts/prompt_v6.txt |
Video_Games_reasoning_v6、Video_Games_reasoning_v6_sample2、Video_Games_reasoning_v6_sample3、Video_Games_reasoning_v6_sample4、Video_Games_reasoning_v6_sample5 |
提示词输出格式
所有提示词均输出包含 predictability、transition_type 和 reasoning 字段的严格 JSON:
- v4:要求 4-8 句分析师独白(证据 → 兴趣 → 机制 → 预测)。
- v5:收紧为 1-6 句基于证据的推理,禁止复述完整历史。
- v6:回归 4-8 句分析师独白(120-180 词),并额外提供物品描述,使每个语义 ID 通过具体属性表征而不提及标题。
模型检查点
数据集中包含已训练模型,存放于 checkpoints/<experiment>/<step>/,格式为 HuggingFace 模型可直接加载:
| 实验名称 | 阶段 | 检查点 |
|---|---|---|
Video_Games_stage2_cot_v4_gpt54_Qwen3-1.7B |
Stage-2 CoT | final_checkpoint |
Video_Games_stage2_cot_v5_gpt54_Qwen3-1.7B |
Stage-2 CoT | final_checkpoint |
Video_Games_stage2_cot_v6_single_Qwen3-1.7B |
Stage-2 CoT | final_checkpoint |
Video_Games_stage2_transition_cot_v6_Qwen3-1.7B |
Stage-2 CoT | final_checkpoint |
Video_Games_stage3_rl_no_kl_cot_v4_gpt54_Qwen3-1.7B |
Stage-3 RL | global_step_200 |
Video_Games_stage3_rl_no_kl_cot_v5_gpt54_Qwen3-1.7B |
Stage-3 RL | global_step_100 |
Stage-3 条目为 GRPO 检查点,基于对应的 Stage-2 CoT 模型初始化,并从分片 FSDP 形式合并生成。
使用方式
加载数据集
python from datasets import load_dataset
ds = load_dataset("mengdanzhu/genrec-dataset", "Video_Games_reasoning_v6_sample5", split="train")
加载模型检查点
python from huggingface_hub import snapshot_download from transformers import AutoModelForCausalLM, AutoTokenizer
sub = "checkpoints/Video_Games_stage3_rl_no_kl_cot_v5_gpt54_Qwen3-1.7B/global_step_100" path = snapshot_download("mengdanzhu/genrec-dataset", repo_type="dataset", allow_patterns=[sub + "/*"]) + "/" + sub model = AutoModelForCausalLM.from_pretrained(path) tokenizer = AutoTokenizer.from_pretrained(path)




