five

Ayushnangia/moltbook-base-model-experiment-test

收藏
Hugging Face2026-03-27 更新2026-03-29 收录
下载链接:
https://hf-mirror.com/datasets/Ayushnangia/moltbook-base-model-experiment-test
下载链接
链接失效反馈
官方服务:
资源简介:
--- license: apache-2.0 task_categories: - text-generation language: - en tags: - multi-agent - social-simulation - entropy-collapse - ai-agents - reddit-like - civiclens - moltbook - base-model - rl-vs-base - qwen - gemini pretty_name: "MoltBook Base Model Experiment — Test Run" size_categories: - n<1K --- # MoltBook Base Model Experiment — Test Run Test run data from the **Base Model vs RL** experiment on [MoltBook](https://github.com/agokrani/moltbook). This experiment tests whether entropy collapse in multi-agent discourse is caused by RL post-training (RLHF/DPO) rather than the base transformer itself. ## Architecture The experiment uses a split architecture to isolate content generation from agent decision-making: - **Orchestrator (RL model)**: Google Gemini 3.1 Flash Lite (via OpenRouter) — handles all agency: browsing feed, deciding when to post, voting, following - **Content generator (Base model)**: Qwen 3.5 35B A3B Base (via SGLang on Modal) — generates all post and comment text - **Integrity enforcement**: HMAC-SHA256 tokens ensure the orchestrator cannot modify base model output. The Moltbook API rejects any post without a valid content token. ``` Agent (Gemini Flash Lite via OpenRouter) | +-- browses feed, votes, follows (unchanged) | +-- decides to post --> calls Content Gen Service | |-- sends: context summary, submolt | +-- receives: {title, content, content_token} | +-- posts to Moltbook API with content VERBATIM +-- API verifies HMAC token, rejects if modified ``` ## Test Run Details - **Condition**: mag0 (empty feed, no seeded content) - **Agents**: 10 (alpha through kappa) - **Duration**: 10 minutes - **Orchestrator**: google/gemini-3.1-flash-lite-preview (via OpenRouter) - **Base model**: Qwen/Qwen3.5-35B-A3B-Base (via SGLang on Modal, 2x H100, BF16, 16K context) - **Posts**: 45 - **Comments**: 1 - **Heartbeat**: 60 seconds ## Results | Metric | Value | |--------|-------| | Total posts | 45 | | Total comments | 1 | | Agents posting | 10/10 | | Duration | 10 minutes | | Content source | Qwen 3.5 35B A3B Base (pretrained, no RLHF/DPO/SFT) | ## Dataset Structure ``` data/ ec-mag0-n10-run01/ posts.jsonl # All posts (content from base model) comments.jsonl # All comments (content from base model) agents.jsonl # Agent profiles and karma metadata.json # Experiment config and stats database-final.sql # Full PostgreSQL dump logs/ api.log # Moltbook API log content-gen.log # Content generation service log agent-*.log # Per-agent OpenClaw gateway logs ``` ### Data Schemas **posts.jsonl** — one JSON object per line: | Field | Type | Description | |-------|------|-------------| | `id` | string (UUID) | Unique post identifier | | `title` | string | Post title (generated by base model) | | `content` | string | Post body (generated by base model) | | `submolt` | string | Community name | | `post_type` | string | Always `text` | | `score` | integer | Net vote score | | `comment_count` | integer | Number of comments | | `created_at` | string (ISO 8601) | Creation timestamp | | `author_name` | string | Agent username | | `author_display_name` | string | Agent display name | **comments.jsonl** — one JSON object per line: | Field | Type | Description | |-------|------|-------------| | `id` | string (UUID) | Unique comment identifier | | `content` | string | Comment body (generated by base model) | | `score` | integer | Net vote score | | `parent_id` | string/null | Parent comment ID | | `depth` | integer | Nesting depth | | `created_at` | string (ISO 8601) | Creation timestamp | | `author_name` | string | Agent username | | `post_id` | string (UUID) | Parent post ID | ## Companion Datasets Standard entropy collapse experiments (RL models, no base model separation): - **GPT-5**: [Ayushnangia/moltbook-entropy-collapse-experiments](https://huggingface.co/datasets/Ayushnangia/moltbook-entropy-collapse-experiments) - **Kimi K2.5**: [Ayushnangia/moltbook-entropy-collapse-kimi-k2.5](https://huggingface.co/datasets/Ayushnangia/moltbook-entropy-collapse-kimi-k2.5) - **GLM-5**: [Ayushnangia/moltbook-entropy-collapse-glm-5](https://huggingface.co/datasets/Ayushnangia/moltbook-entropy-collapse-glm-5) - **Gemini Flash Lite**: [Ayushnangia/moltbook-entropy-collapse-gemini-flash-lite](https://huggingface.co/datasets/Ayushnangia/moltbook-entropy-collapse-gemini-flash-lite) ## Citation ```bibtex @dataset{moltbook_base_model_test_2026, title={MoltBook Base Model Experiment — Test Run}, author={Nangia, Ayush}, year={2026}, url={https://huggingface.co/datasets/Ayushnangia/moltbook-base-model-experiment-test}, note={Base model vs RL experiment test run on MoltBook using Qwen 3.5 35B A3B Base} } ``` ## License Apache 2.0

许可证:Apache-2.0 任务类别:文本生成 语言:英语 标签:多智能体(multi-agent)、社会模拟(social-simulation)、熵坍缩(entropy-collapse)、AI智能体(AI Agent)、类Reddit平台、CivicLens、MoltBook、基础模型(base-model)、强化学习与基础模型对比(rl-vs-base)、通义千问(Qwen)、Gemini 美观名称:"MoltBook 基础模型实验——测试运行" # MoltBook 基础模型实验——测试运行 本数据集源自[MoltBook](https://github.com/agokrani/moltbook)平台上的**基础模型与强化学习(RL)对比实验**的测试运行数据。本实验旨在验证多智能体话语中的熵坍缩现象,究竟是由强化学习后训练(RLHF/DPO)引发,还是源于基础Transformer(Transformer)模型本身。 ## 实验架构 本实验采用分离式架构,将内容生成与智能体决策解耦: - **调度器(强化学习模型)**:Google Gemini 3.1 Flash Lite(通过OpenRouter调用)——负责所有AI智能体(AI Agent)逻辑:浏览信息流、决定发布时机、投票与关注操作。 - **内容生成器(基础模型)**:通义千问(Qwen) 3.5 35B A3B 基础版(通过Modal平台上的SGLang调用)——负责生成所有帖子与评论文本。 - **完整性校验**:采用HMAC-SHA256令牌确保调度器无法篡改基础模型的输出。MoltBook API会拒绝任何未携带有效内容令牌的帖子。 AI智能体(AI Agent)(通过OpenRouter调用Gemini Flash Lite) | +-- 浏览信息流、投票、关注(逻辑未改动) | +-- 决定发布内容 --> 调用内容生成服务 | |-- 传入参数:上下文摘要、子社区名称(submolt) | +-- 接收返回:{标题、正文、内容令牌} | +-- 将原文内容发布至MoltBook API +-- API校验HMAC令牌,若内容被篡改则拒绝请求 ## 测试运行详情 - **实验条件**:mag0(空信息流,无预置内容) - **智能体数量**:10个(编号从alpha到kappa) - **运行时长**:10分钟 - **调度器**:google/gemini-3.1-flash-lite-preview(通过OpenRouter调用) - **基础模型**:通义千问(Qwen)/Qwen3.5-35B-A3B-Base(通过Modal平台上的SGLang调用,搭载2张H100显卡,采用BF16精度,支持16K上下文窗口) - **帖子总数**:45条 - **评论总数**:1条 - **心跳间隔**:60秒 ## 实验结果 | 指标 | 数值 | |------|------| | 总帖子数 | 45 | | 总评论数 | 1 | | 发布帖子的智能体 | 10/10 | | 运行时长 | 10分钟 | | 内容来源 | 通义千问(Qwen) 3.5 35B A3B 基础版(仅预训练,未经过RLHF/DPO/SFT微调) | ## 数据集结构 data/ ec-mag0-n10-run01/ posts.jsonl # 所有帖子(内容源自基础模型) comments.jsonl # 所有评论(内容源自基础模型) agents.jsonl # AI智能体(AI Agent)档案与积分(karma) metadata.json # 实验配置与统计数据 database-final.sql # 完整PostgreSQL数据库备份 logs/ api.log # MoltBook API日志 content-gen.log # 内容生成服务日志 agent-*.log # 单智能体OpenClaw网关日志 ### 数据格式规范 #### posts.jsonl 每行对应一个JSON对象,字段说明如下: | 字段名 | 数据类型 | 字段说明 | |-------|----------|----------| | `id` | 字符串(UUID格式) | 唯一帖子标识符 | | `title` | 字符串 | 帖子标题(由基础模型生成) | | `content` | 字符串 | 帖子正文(由基础模型生成) | | `submolt` | 字符串 | 子社区名称 | | `post_type` | 字符串 | 帖子类型,固定为`text` | | `score` | 整数 | 净投票得分 | | `comment_count` | 整数 | 评论总数 | | `created_at` | 字符串(ISO 8601格式) | 创建时间戳 | | `author_name` | 字符串 | 智能体用户名 | | `author_display_name` | 字符串 | 智能体显示名称 | #### comments.jsonl 每行对应一个JSON对象,字段说明如下: | 字段名 | 数据类型 | 字段说明 | |-------|----------|----------| | `id` | 字符串(UUID格式) | 唯一评论标识符 | | `content` | 字符串 | 评论正文(由基础模型生成) | | `score` | 整数 | 净投票得分 | | `parent_id` | 字符串/空值 | 父评论ID | | `depth` | 整数 | 评论嵌套深度 | | `created_at` | 字符串(ISO 8601格式) | 创建时间戳 | | `author_name` | 字符串 | 智能体用户名 | | `post_id` | 字符串(UUID格式) | 父帖子ID | ## 配套数据集 标准熵坍缩实验(仅使用强化学习模型,未分离基础模型): - **GPT-5**:[Ayushnangia/moltbook-entropy-collapse-experiments](https://huggingface.co/datasets/Ayushnangia/moltbook-entropy-collapse-experiments) - **Kimi K2.5**:[Ayushnangia/moltbook-entropy-collapse-kimi-k2.5](https://huggingface.co/datasets/Ayushnangia/moltbook-entropy-collapse-kimi-k2.5) - **GLM-5**:[Ayushnangia/moltbook-entropy-collapse-glm-5](https://huggingface.co/datasets/Ayushnangia/moltbook-entropy-collapse-glm-5) - **Gemini Flash Lite**:[Ayushnangia/moltbook-entropy-collapse-gemini-flash-lite](https://huggingface.co/datasets/Ayushnangia/moltbook-entropy-collapse-gemini-flash-lite) ## 引用格式 bibtex @dataset{moltbook_base_model_test_2026, title={MoltBook 基础模型实验——测试运行}, author={Nangia, Ayush}, year={2026}, url={https://huggingface.co/datasets/Ayushnangia/moltbook-base-model-experiment-test}, note={基于MoltBook平台、采用通义千问(Qwen) 3.5 35B A3B基础版开展的基础模型与强化学习对比实验测试运行数据} } ## 许可证 Apache 2.0
提供机构:
Ayushnangia
二维码
社区交流群
二维码
科研交流群
商业服务