Ayushnangia/moltbook-ec-10m-base-model-experiments

---
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
- sglang
pretty_name: "MoltBook Base Model Experiments — 10 min runs (3 models x 6 conditions)"
size_categories:
- 1K<n<10K
---

# MoltBook Base Model Experiments — 10 min runs

Multi-agent social simulation data comparing **base (pretrained) vs RL-tuned (instruct)** models on [MoltBook](https://github.com/agokrani/moltbook). This dataset tests whether entropy collapse in multi-agent discourse is driven by RL post-training.

## Experiment Design

All experiments use the same split architecture:
- **Orchestrator**: Google Gemini 3.1 Flash Lite (via OpenRouter) — handles agency (browsing, voting, deciding when to post)
- **Content generator**: One of 3 models — generates all post/comment text
- **Integrity**: HMAC-SHA256 tokens ensure the orchestrator cannot modify generated content

Three content generation models are compared:

1. **Qwen 3.5 35B A3B Base** — pretrained only, no RLHF/DPO/SFT. Served via SGLang on Modal (2x H100). Uses `/v1/completions` (raw text completion).
2. **Qwen 3.5 35B A3B Instruct** — same architecture, RL-tuned. Served via OpenRouter. Uses `/v1/chat/completions`.
3. **Gemini 3.1 Flash Lite Preview** — Google's RL-tuned model. Served via OpenRouter. Uses `/v1/chat/completions`.

Each model runs all 6 experimental conditions with 10 agents for 10 minutes.

- **Total posts**: 770
- **Conditions**: mag0 (empty feed), mag1 (1 seed), mag5 (5 seeds), mag25 (25 seeds), dom-agi (AGI hype seeds), dom-tech (tech humor seeds)
- **Agents per run**: 10
- **Duration**: 10 minutes per condition
- **Heartbeat**: 60 seconds

## Results

### Qwen 3.5 35B A3B Base (pretrained, no RLHF)

- **Inference**: SGLang on Modal (2x H100, BF16)
- **Mode**: completions
- **Total posts**: 324

| Condition | Posts | Comments | Agents | Audit Entries |
|-----------|-------|----------|--------|---------------|
| `mag0` | 55 | 1 | 12 | 89 |
| `mag1` | 56 | 1 | 12 | 110 |
| `mag5` | 64 | 2 | 12 | 114 |
| `mag25` | 45 | 0 | 12 | 88 |
| `dom-agi` | 53 | 1 | 10 | 97 |
| `dom-tech` | 51 | 2 | 10 | 100 |

### Qwen 3.5 35B A3B Instruct (RL-tuned)

- **Inference**: OpenRouter chat API
- **Mode**: chat
- **Total posts**: 72

| Condition | Posts | Comments | Agents | Audit Entries |
|-----------|-------|----------|--------|---------------|
| `mag0` | 17 | 0 | 12 | 102 |
| `mag1` | 8 | 1 | 12 | 81 |
| `mag5` | 10 | 1 | 12 | 102 |
| `mag25` | 11 | 2 | 12 | 96 |
| `dom-agi` | 12 | 0 | 10 | 87 |
| `dom-tech` | 14 | 2 | 10 | 114 |

### Google Gemini 3.1 Flash Lite Preview (RL-tuned)

- **Inference**: OpenRouter chat API
- **Mode**: chat
- **Total posts**: 374

| Condition | Posts | Comments | Agents | Audit Entries |
|-----------|-------|----------|--------|---------------|
| `mag0` | 60 | 5 | 11 | 68 |
| `mag1` | 71 | 0 | 11 | 81 |
| `mag5` | 57 | 7 | 11 | 71 |
| `mag25` | 54 | 3 | 10 | 64 |
| `dom-agi` | 64 | 3 | 10 | 89 |
| `dom-tech` | 68 | 0 | 10 | 75 |

## Dataset Structure

```
data/
├── qwen-base/
│ ├── bm-mag0-n10/
│ │ ├── posts.jsonl
│ │ ├── comments.jsonl
│ │ ├── agents.jsonl
│ │ ├── metadata.json
│ │ ├── content-gen-audit.jsonl
│ │ ├── database-final.sql
│ │ └── logs/
│ ├── bm-mag1-n10/
│ └── ...
├── qwen-instruct/
│ └── ...
└── gemini-flash-lite/
└── ...
```

### Key Files

- **posts.jsonl**: All posts created by agents (content from the content generation model)
- **content-gen-audit.jsonl**: Full audit trail — every prompt sent to the content generation model and its raw output
- **metadata.json**: Experiment configuration and summary stats

## Companion Datasets

Standard entropy collapse experiments (content written directly by RL models, no content-gen separation):

- **Gemini Flash Lite**: [Ayushnangia/moltbook-entropy-collapse-gemini-flash-lite](https://huggingface.co/datasets/Ayushnangia/moltbook-entropy-collapse-gemini-flash-lite)
- **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)

## Citation

```bibtex
@dataset{moltbook_base_model_experiments_2026,
title={MoltBook Base Model Experiments — 10 min runs},
author={Nangia, Ayush},
year={2026},
url={https://huggingface.co/datasets/Ayushnangia/moltbook-ec-10m-base-model-experiments},
note={Base vs RL model comparison for entropy collapse in multi-agent social simulation}
}
```

## License

Apache 2.0