withmartian/ares-20q-case-study

---
license: apache-2.0
tags:
- mechanistic-interpretability
- steering-vectors
- transformer-lens
- agents
- reinforcement-learning
---

# ARES 20Q Case Study — Pre-computed Results

Pre-computed activations and experiment results for the
[ARES Mechanistic Interpretability tutorial](https://github.com/withmartian/ares/blob/main/examples/20q_case_study/ares_mi_20q_tutorial.ipynb).

## What's included

| Directory | Size | Description |
|-----------|------|-------------|
| `20q_data/` | ~4.6 GB | 50 episodes of Llama-3.2-1B-Instruct playing Twenty Questions, with layer-8 residual stream activations captured at every step |
| `20q_probing_results/` | 137 KB | Linear probe accuracy plots (per-step and global) |
| `20q_steering_vector_evolution/` | 417 KB | Cosine similarity heatmap, PCA trajectory, and norm plots for per-step steering vectors |
| `deterministic_20q_steering_results/` | 881 KB | Steering experiment results (baseline + 4 alpha values, 20 episodes each) |

## Quick start

```python
from huggingface_hub import snapshot_download

# Download everything (~4.6 GB)
snapshot_download(
repo_id="withmartian/ares-20q-case-study",
repo_type="dataset",
local_dir="outputs",
)

# Or download only the lightweight analysis results (~1.5 MB)
snapshot_download(
repo_id="withmartian/ares-20q-case-study",
repo_type="dataset",
local_dir="outputs",
ignore_patterns=["20q_data/*"],
)
```

## Model & setup

- **Model**: `meta-llama/Llama-3.2-1B-Instruct` (via TransformerLens)
- **Hook point**: `blocks.8.hook_resid_post` (middle layer residual stream)
- **Episodes**: 50 (data collection), 20 per condition (steering)
- **Max steps per episode**: 25 (data collection), 20 (steering)

## Citation

If you use this data, please cite the ARES repository:

```
@software{ares2025,
title={ARES: Agentic Research and Evaluation Suite},
url={https://github.com/withmartian/ares},
year={2025}
}
```

许可证：Apache-2.0
标签：
- 机制可解释性（mechanistic interpretability）
- 转向向量（steering vector）
- Transformer Lens（transformer-lens）
- AI 智能体（agents）
- 强化学习（reinforcement-learning）

# ARES 20Q案例研究——预计算结果

本数据集为[ARES机制可解释性教程](https://github.com/withmartian/ares/blob/main/examples/20q_case_study/ares_mi_20q_tutorial.ipynb)提供预计算的激活值与实验结果。

## 包含内容

| 目录路径 | 大小 | 描述 |
|-----------|------|-------------|
| `20q_data/` | 约4.6 GB | 包含50轮Llama-3.2-1B-Instruct模型玩二十问（Twenty Questions）的实验数据，每一步均捕获了第8层残差流（residual stream）的激活值 |
| `20q_probing_results/` | 137 KB | 线性探测（linear probe）的准确率绘图结果（包含逐步与全局两种形式） |
| `20q_steering_vector_evolution/` | 417 KB | 逐步转向向量的余弦相似度热图、PCA轨迹与范数绘图结果 |
| `deterministic_20q_steering_results/` | 881 KB | 转向实验的结果（包含基线组与4组不同α值的实验组，每组各20轮实验） |

## 快速上手

python
from huggingface_hub import snapshot_download

# 下载全部数据集（约4.6 GB）
snapshot_download(
repo_id="withmartian/ares-20q-case-study",
repo_type="dataset",
local_dir="outputs",
)

# 或仅下载轻量型分析结果（约1.5 MB）
snapshot_download(
repo_id="withmartian/ares-20q-case-study",
repo_type="dataset",
local_dir="outputs",
ignore_patterns=["20q_data/*"],
)

## 模型与实验设置

- **模型**：`meta-llama/Llama-3.2-1B-Instruct`（通过TransformerLens加载）
- **挂钩点**：`blocks.8.hook_resid_post`（中间层残差流）
- **实验轮次**：数据收集阶段共50轮，转向实验阶段每组20轮
- **单轮最大步数**：数据收集阶段为25步，转向实验阶段为20步

## 引用

若使用本数据集，请引用ARES仓库：

@software{ares2025,
title={ARES: Agentic Research and Evaluation Suite},
url={https://github.com/withmartian/ares},
year={2025}
}