| Chapter 0: Fundamentals |
| 0.0Prerequisites |
โฌ |
โ |
~1 min |
๐ข |
โฌ Under 10 MB |
<1 GB |
None |
- ๐ฅ๏ธ GPU: โฌ Not needed.
- โฑ๏ธ Runtime (ex. downloads/API): CPU ~1 min
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: โฌ Under 10 MB
- ๐ API keys: None
- ๐ RAM: <1 GB at peak
|
| 0.1Ray Tracing |
๐ฉ 3.6 GiB |
~2.2 min |
~3 min (CUDA cells skipped) |
๐ข |
โฌ Under 10 MB |
~3.7 GB |
None |
- ๐ฅ๏ธ GPU: ๐ฉ Optional, peak ~3.6 GiB. Video & lighting cells need CUDA.
- โฑ๏ธ Runtime (ex. downloads/API): GPU ~2.2 min ยท CPU ~3 min (CUDA cells skipped)
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: โฌ Under 10 MB
- ๐ API keys: None
- ๐ RAM: ~3.7 GB at peak
|
| 0.2CNNs & ResNets |
๐ฉ 0.4 GiB* |
<1 min |
~5 min (bonus ~36 min/epoch) |
๐ข |
โฌ ~150 MB |
~1.4 GB |
None |
- ๐ฅ๏ธ GPU: ๐ฉ Optional, peak ~0.4 GiB.
- ๐จ Recommended, Bonus - Feature extraction: ~36 min/epoch on CPU
- โฑ๏ธ Runtime (ex. downloads/API): GPU <1 min ยท CPU ~5 min (bonus ~36 min/epoch)
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: โฌ ~150 MB
- ๐ API keys: None
- ๐ RAM: ~1.4 GB at peak
|
| 0.3Optimization |
๐จ 4.4 GiB* |
~2 min |
~hours |
๐ก |
โฌ ~230 MB |
~2.2 GB |
Weights & Biases |
- ๐ฅ๏ธ GPU: ๐จ Recommended, peak ~4.4 GiB.
- ๐ฅ Big GPU, Distributed Training: 2+ GPUs
- โฑ๏ธ Runtime (ex. downloads/API): GPU ~2 min ยท CPU ~hours
- โ๏ธ Colab? ๐ก Partly: no Distributed Training (needs 2+ GPUs)
- ๐พ Downloads: โฌ ~230 MB
- ๐ API keys:
- ๐ฅ Mandatory: Weights & Biases, for the hyperparameter sweeps
- ๐ RAM: ~2.2 GB at peak
|
| 0.4Backpropagation |
โฌ |
โ |
~1 min |
๐ข |
โฌ ~20 MB |
~1 GB |
None |
- ๐ฅ๏ธ GPU: โฌ Not needed.
- โฑ๏ธ Runtime (ex. downloads/API): CPU ~1 min
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: โฌ ~20 MB
- ๐ API keys: None
- ๐ RAM: ~1 GB at peak
|
| 0.5.1VAEs |
๐ฉ 0.5 GiB |
<1 min |
~7 min |
๐ข |
โฌ ~20 MB |
~1.5 GB |
Weights & Biases (optional) |
- ๐ฅ๏ธ GPU: ๐ฉ Optional, peak ~0.5 GiB.
- โฑ๏ธ Runtime (ex. downloads/API): GPU <1 min ยท CPU ~7 min
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: โฌ ~20 MB
- ๐ API keys:
- ๐ฉ Optional: Weights & Biases, for logging (
use_wandb=True by default; set it to False to skip)
- ๐ RAM: ~1.5 GB at peak
|
| 0.5.2GANs |
๐จ 1.6 GiB |
~6.9 min |
~hours |
๐ข |
๐ฉ ~1.4 GB |
~4.4 GB |
Weights & Biases (optional) |
- ๐ฅ๏ธ GPU: ๐จ Recommended, peak ~1.6 GiB.
- โฑ๏ธ Runtime (ex. downloads/API): GPU ~6.9 min ยท CPU ~hours
- โ๏ธ Colab? ๐ข Yes: tested under T4 GPU and RAM limits
- ๐พ Downloads: ๐ฉ ~1.4 GB: nielsr/CelebA-faces 1.4 GB
- ๐ API keys:
- ๐ฉ Optional: Weights & Biases, for logging training runs (
use_wandb)
- ๐ RAM: ~4.4 GB at peak
|
| Chapter 1: Transformer Interpretability |
| 1.1Transformers from Scratch |
๐จ 6.6 GiB |
~11 min |
~3 h |
๐ข |
๐ฉ ~800 MB |
~3.5 GB |
None |
- ๐ฅ๏ธ GPU: ๐จ Recommended, peak ~6.6 GiB.
- โฑ๏ธ Runtime (ex. downloads/API): GPU ~11 min ยท CPU ~3 h
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: ๐ฉ ~800 MB: gpt2 0.55 GB, roneneldan/TinyStories 0.24 GB
- ๐ API keys: None
- ๐ RAM: ~3.5 GB at peak
|
| 1.2Intro to Mech Interp |
๐ฉ 6.2 GiB |
<1 min |
~3 min |
๐ข |
๐ฉ ~740 MB |
~1.8 GB |
None |
- ๐ฅ๏ธ GPU: ๐ฉ Optional, peak ~6.2 GiB.
- โฑ๏ธ Runtime (ex. downloads/API): GPU <1 min ยท CPU ~3 min
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: ๐ฉ ~740 MB: gpt2 0.55 GB
- ๐ API keys: None
- ๐ RAM: ~1.8 GB at peak
|
| 1.3.1Linear Probes |
๐ง 27 GiB (14 GiB) |
~18 min |
N/A |
โ |
๐ง ~86 GB |
~13 GB |
Hugging Face |
- ๐ฅ๏ธ GPU: ๐ง 16 GiB (A4000). Peak ~14 GiB with
LOW_GPU_MEM (auto below 30 GiB: 13B and 8B in 8-bit, 70B bonus in 4-bit); ~27 GiB without.
- โฑ๏ธ Runtime (ex. downloads/API): GPU ~18 min ยท CPU N/A
- โ๏ธ Colab? โ Untested: RAM ~13 GB vs 12.7; 86 GB downloads vs ~78 GB disk; bf16 on a T4
- ๐พ Downloads: ๐ง ~86 GB: meta-llama/Llama-3.3-70B-Instruct (first 24 of 80 layers, for the 70B section) 43 GB, meta-llama/Llama-2-13b-hf 26 GB, meta-llama/Meta-Llama-3.1-8B-Instruct 16 GB
- ๐ API keys:
- ๐ฅ Mandatory: Hugging Face (
HF_TOKEN), for the Llama models (request access: meta-llama/Llama-2-13b-hf, meta-llama/Llama-3.1-8B-Instruct, meta-llama/Llama-3.3-70B-Instruct (bonus))
- ๐ RAM: ~13 GB at peak
|
| 1.3.2Function Vectors & Model Steering |
๐ง 22 GiB (14 GiB) |
~1.3 min |
N/A (hours, ~34 GB RAM) |
โ |
๐ง ~31 GB |
~3.7 GB |
None |
- ๐ฅ๏ธ GPU: ๐ง 16 GiB (A4000). Peak ~14 GiB with
LOW_GPU_MEM (auto below 23 GiB: sweeps one layer at a time, same results); ~22 GiB without.
- โฑ๏ธ Runtime (ex. downloads/API): GPU ~1.3 min ยท CPU N/A (hours, ~34 GB RAM)
- โ๏ธ Colab? โ Untested: loading GPT-J's 24 GB fp32 checkpoint in 12.7 GB RAM
- ๐พ Downloads: ๐ง ~31 GB: EleutherAI/gpt-j-6b 24 GB, gpt2-xl 6.4 GB
- ๐ API keys: None
- ๐ RAM: ~3.7 GB at peak
|
| 1.3.3Interpretability with SAEs |
๐ง 30 GiB (13 GiB) |
~9.3 min |
N/A (Gemma needs ~27 GB RAM) |
โ |
๐จ ~15 GB |
~12 GB |
Hugging Face, OpenAI (optional), Weights & Biases (optional) |
- ๐ฅ๏ธ GPU: ๐ง 16 GiB (A4000). Peak ~13 GiB with
LOW_GPU_MEM (auto below 28 GiB: gemma-2b-it in bf16, smaller sae_vis batches); ~30 GiB without.
- โฑ๏ธ Runtime (ex. downloads/API): GPU ~9.3 min ยท CPU N/A (Gemma needs ~27 GB RAM)
- โ๏ธ Colab? โ Untested: RAM ~12 GB vs 12.7; bf16 on a T4
- ๐พ Downloads: ๐จ ~15 GB: google/gemma-2-2b 5.2 GB, google/gemma-2b-it 5.0 GB, ckkissane/attn-saes-gpt2-small-all-layers 2.1 GB, jbloom/GPT2-Small-Feature-Splitting-Experiment-Layer-8 0.6 GB
- ๐ API keys:
- ๐ฅ Mandatory: Hugging Face (
HF_TOKEN), for the Gemma models (request access: google/gemma-2-2b, google/gemma-2b-it)
- ๐ฉ Optional: OpenAI (
OPENAI_API_KEY), for autointerp
- ๐ฉ Optional: Weights & Biases, for the SAE-training section
- ๐ RAM: ~12 GB at peak
|
| 1.3.4Activation Oracles |
๐ง 19 GiB (13 GiB) |
~8.4 min |
~hours (~24 GB RAM) |
โ |
๐ง ~37 GB |
~13 GB |
Hugging Face |
- ๐ฅ๏ธ GPU: ๐ง 16 GiB (A4000). Peak ~13 GiB with
LOW_GPU_MEM (auto below 20 GiB: both 8B models in 8-bit); ~19 GiB without.
- โฑ๏ธ Runtime (ex. downloads/API): GPU ~8.4 min ยท CPU ~hours (~24 GB RAM)
- โ๏ธ Colab? โ Untested: RAM ~13 GB vs 12.7
- ๐พ Downloads: ๐ง ~37 GB: Qwen/Qwen3-8B 16 GB, meta-llama/Meta-Llama-3.1-8B-Instruct 16 GB, adamkarvonen/checkpoints_latentqa_cls_past_lens_addition_Qwen3-8B 0.7 GB, adamkarvonen/checkpoints_latentqa_cls_past_lens_Llama-3_1-8B-Instruct 0.67 GB
- ๐ API keys:
- ๐ฅ Mandatory: Hugging Face (
HF_TOKEN), for the model-diffing section (request access: meta-llama/Llama-3.1-8B-Instruct)
- ๐ RAM: ~13 GB at peak
|
| 1.4.1Indirect Object Identification |
๐ฉ 2.8 GiB |
~1 min |
~17 min |
๐ข |
๐ฉ ~550 MB |
~2 GB |
None |
- ๐ฅ๏ธ GPU: ๐ฉ Optional, peak ~2.8 GiB.
- โฑ๏ธ Runtime (ex. downloads/API): GPU ~1 min ยท CPU ~17 min
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: ๐ฉ ~550 MB: gpt2 0.55 GB
- ๐ API keys: None
- ๐ RAM: ~2 GB at peak
|
| 1.4.2SAE Circuits |
๐ง 27 GiB (14 GiB) |
~5.6 min |
N/A (~40 GB RAM) |
โ |
๐ง ~46 GB |
~15 GB |
Hugging Face |
- ๐ฅ๏ธ GPU: ๐ง 16 GiB (A4000). Peak ~14 GiB with
LOW_GPU_MEM (auto below 28 GiB: smaller batches, circuit-tracer in bf16); ~27 GiB without.
- โฑ๏ธ Runtime (ex. downloads/API): GPU ~5.6 min ยท CPU N/A (~40 GB RAM)
- โ๏ธ Colab? โ Untested: RAM ~15 GB vs 12.7; bf16 on a T4
- ๐พ Downloads: ๐ง ~46 GB: google/gemma-scope-2-1b-it 25 GB, mwhanna/gemma-scope-transcoders 7.9 GB, google/gemma-2-2b 5.2 GB, ckkissane/attn-saes-gpt2-small-all-layers 2.1 GB
- ๐ API keys:
- ๐ฅ Mandatory: Hugging Face (
HF_TOKEN), for the Gemma models (request access: google/gemma-2-2b, google/gemma-3-1b-it)
- ๐ RAM: ~15 GB at peak
|
| 1.5.1Balanced Bracket Classifier |
๐ฉ 5.8 GiB |
<1 min |
~1 min |
๐ข |
โฌ Under 10 MB |
~1.5 GB |
None |
- ๐ฅ๏ธ GPU: ๐ฉ Optional, peak ~5.8 GiB.
- โฑ๏ธ Runtime (ex. downloads/API): GPU <1 min ยท CPU ~1 min
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: โฌ Under 10 MB
- ๐ API keys: None
- ๐ RAM: ~1.5 GB at peak
|
| 1.5.2Grokking & Modular Arithmetic |
๐ฉ 1.6 GiB |
<1 min |
~5 min |
๐ข |
๐ฉ ~460 MB |
~1.5 GB |
None |
- ๐ฅ๏ธ GPU: ๐ฉ Optional, peak ~1.6 GiB.
- โฑ๏ธ Runtime (ex. downloads/API): GPU <1 min ยท CPU ~5 min
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: ๐ฉ ~460 MB
- ๐ API keys: None
- ๐ RAM: ~1.5 GB at peak
|
| 1.5.3OthelloGPT |
๐จ 1.5 GiB |
<1 min |
~5 min |
๐ข |
โฌ ~100 MB |
~1.6 GB |
Weights & Biases (optional) |
- ๐ฅ๏ธ GPU: ๐จ Recommended, peak ~1.5 GiB.
- โฑ๏ธ Runtime (ex. downloads/API): GPU <1 min ยท CPU ~5 min
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: โฌ ~100 MB
- ๐ API keys:
- ๐ฉ Optional: Weights & Biases, for logging training runs (
use_wandb)
- ๐ RAM: ~1.6 GB at peak
|
| 1.5.4Superposition & SAEs |
๐จ 1.9 GiB |
~11 min |
~40 min |
๐ข |
โฌ ~30 MB |
~2.3 GB |
None |
- ๐ฅ๏ธ GPU: ๐จ Recommended, peak ~1.9 GiB.
- โฑ๏ธ Runtime (ex. downloads/API): GPU ~11 min ยท CPU ~40 min
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: โฌ ~30 MB
- ๐ API keys: None
- ๐ RAM: ~2.3 GB at peak
|
| Chapter 2: Reinforcement Learning |
| 2.1Intro to RL |
โฌ |
โ |
~1 min |
๐ข |
โฌ Under 10 MB |
<1 GB |
None |
- ๐ฅ๏ธ GPU: โฌ Not needed.
- โฑ๏ธ Runtime (ex. downloads/API): CPU ~1 min
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: โฌ Under 10 MB
- ๐ API keys: None
- ๐ RAM: <1 GB at peak
|
| 2.2.1DQN |
โฌ |
โ |
~1 min |
๐ข |
โฌ Under 10 MB |
~1.6 GB |
Weights & Biases |
- ๐ฅ๏ธ GPU: โฌ Not needed. CPU is faster.
- โฑ๏ธ Runtime (ex. downloads/API): CPU ~1 min
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: โฌ Under 10 MB
- ๐ API keys:
- ๐ฅ Mandatory: Weights & Biases, for the training runs that set
use_wandb=True (or set WANDB_MODE=offline)
- ๐ RAM: ~1.6 GB at peak
|
| 2.2.2VPG |
โฌ |
โ |
<1 min |
๐ข |
โฌ Under 10 MB |
~1.5 GB |
Weights & Biases (optional) |
- ๐ฅ๏ธ GPU: โฌ Not needed. CPU is faster.
- โฑ๏ธ Runtime (ex. downloads/API): CPU <1 min
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: โฌ Under 10 MB
- ๐ API keys:
- ๐ฉ Optional: Weights & Biases, for logging training runs (
use_wandb)
- ๐ RAM: ~1.5 GB at peak
|
| 2.3PPO |
๐ฉ |
~1.3 min |
minutes for CartPole; Atari & MuJoCo much longer |
โ |
โฌ Under 10 MB |
~2.3 GB |
Weights & Biases (optional) |
- ๐ฅ๏ธ GPU: ๐ฉ Optional.
- โฑ๏ธ Runtime (ex. downloads/API): GPU ~1.3 min ยท CPU minutes for CartPole; Atari & MuJoCo much longer
- โ๏ธ Colab? โ Untested: MuJoCo (jax/brax install, video rendering) not tried
- ๐พ Downloads: โฌ Under 10 MB
- ๐ API keys:
- ๐ฉ Optional: Weights & Biases, for logging training runs (
use_wandb)
- ๐ RAM: ~2.3 GB at peak
- ๐ Also: Headless MuJoCo videos need OSMesa (
libosmesa6 or conda mesalib).
|
| 2.3.2Goal Misgeneralisation |
๐ฉ 0.9 GiB |
<1 min |
~13 min |
๐ข |
โฌ Under 10 MB |
~2.8 GB |
None |
- ๐ฅ๏ธ GPU: ๐ฉ Optional, peak ~0.9 GiB.
- โฑ๏ธ Runtime (ex. downloads/API): GPU <1 min ยท CPU ~13 min
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: โฌ Under 10 MB
- ๐ API keys: None
- ๐ RAM: ~2.8 GB at peak
|
| 2.4.1RLHF |
๐จ 13 GiB (5.4 GiB) |
~2.5 min |
~hours |
๐ข |
๐จ ~2 GB |
~2.6 GB |
None |
- ๐ฅ๏ธ GPU: ๐จ Recommended. Peak ~5.4 GiB with
LOW_GPU_MEM (auto below 22 GiB: gpt2-small instead of gpt2-medium); ~13 GiB without.
- โฑ๏ธ Runtime (ex. downloads/API): GPU ~2.5 min ยท CPU ~hours
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: ๐จ ~2 GB: Qwen/Qwen3-0.6B-Base 1.2 GB, gpt2 0.55 GB
- ๐ API keys: None
- ๐ RAM: ~2.6 GB at peak
|
| 2.4.2RLVR |
๐ง 16 GiB (9.9 GiB) |
~7.2 min |
N/A |
โ |
๐ฉ ~1.2 GB |
~2.6 GB |
None |
- ๐ฅ๏ธ GPU: ๐ง 16 GiB (A4000). Peak ~9.9 GiB with
LOW_GPU_MEM (auto below 16 GiB: micro-batches of 4, no vLLM on a T4); ~16 GiB without.
- โฑ๏ธ Runtime (ex. downloads/API): GPU ~7.2 min ยท CPU N/A
- โ๏ธ Colab? โ Untested: fits the T4 (~10 GiB), not run on one
- ๐พ Downloads: ๐ฉ ~1.2 GB: Qwen/Qwen3-0.6B-Base 1.2 GB
- ๐ API keys: None
- ๐ RAM: ~2.6 GB at peak
|
| 2.5MCTS & AlphaZero |
๐จ 3.2 GiB |
~2.5 min |
~hours |
๐ข |
โฌ Under 10 MB |
~1.9 GB |
None |
- ๐ฅ๏ธ GPU: ๐จ Recommended, peak ~3.2 GiB.
- โฑ๏ธ Runtime (ex. downloads/API): GPU ~2.5 min ยท CPU ~hours
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: โฌ Under 10 MB
- ๐ API keys: None
- ๐ RAM: ~1.9 GB at peak
|
| Chapter 3: LLM Evaluations |
| 3.1Intro to Evals |
โฌ |
โ |
<1 min |
๐ข |
โฌ Under 10 MB |
<1 GB |
OpenRouter |
- ๐ฅ๏ธ GPU: โฌ Not needed.
- โฑ๏ธ Runtime (ex. downloads/API): CPU <1 min
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: โฌ Under 10 MB
- ๐ API keys:
- ๐ฅ Mandatory: OpenRouter (
OPENROUTER_API_KEY), for the LLM API calls throughout the day
- ๐ RAM: <1 GB at peak
|
| 3.2Dataset Generation |
โฌ |
โ |
<1 min |
๐ข |
โฌ Under 10 MB |
<1 GB |
OpenRouter |
- ๐ฅ๏ธ GPU: โฌ Not needed.
- โฑ๏ธ Runtime (ex. downloads/API): CPU <1 min
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: โฌ Under 10 MB
- ๐ API keys:
- ๐ฅ Mandatory: OpenRouter (
OPENROUTER_API_KEY), for the LLM API calls throughout the day
- ๐ RAM: <1 GB at peak
|
| 3.3Running Evals with Inspect |
โฌ |
โ |
<1 min |
๐ข |
โฌ Under 10 MB |
<1 GB |
OpenRouter |
- ๐ฅ๏ธ GPU: โฌ Not needed.
- โฑ๏ธ Runtime (ex. downloads/API): CPU <1 min
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: โฌ Under 10 MB
- ๐ API keys:
- ๐ฅ Mandatory: OpenRouter (
OPENROUTER_API_KEY), for the LLM API calls throughout the day
- ๐ RAM: <1 GB at peak
|
| 3.4LLM Agents |
โฌ |
โ |
<1 min |
๐ข |
โฌ Under 10 MB |
<1 GB |
OpenRouter |
- ๐ฅ๏ธ GPU: โฌ Not needed.
- โฑ๏ธ Runtime (ex. downloads/API): CPU <1 min
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: โฌ Under 10 MB
- ๐ API keys:
- ๐ฅ Mandatory: OpenRouter (
OPENROUTER_API_KEY), for the LLM API calls throughout the day
- ๐ RAM: <1 GB at peak
|
| 3.5AI Control |
โฌ |
โ |
<1 min |
๐ก |
โฌ ~20 MB |
<1 GB |
OpenRouter |
- ๐ฅ๏ธ GPU: โฌ Not needed.
- โฑ๏ธ Runtime (ex. downloads/API): CPU <1 min
- โ๏ธ Colab? ๐ก Partly: no Docker: set
USING_DOCKER = False
- ๐พ Downloads: โฌ ~20 MB
- ๐ API keys:
- ๐ฅ Mandatory: OpenRouter (
OPENROUTER_API_KEY), for the LLM API calls throughout the day
- ๐ RAM: <1 GB at peak
- ๐ Also: Docker for sandboxing (or set
USING_DOCKER = False).
|
| Chapter 4: Alignment Science |
| 4.1Emergent Misalignment |
๐ง 30 GiB (12 GiB) |
~9.4 min (+ API calls) |
N/A |
โ |
๐ง ~30 GB (~11 GB with LOW_GPU_MEM) |
~6.7 GB |
OpenRouter, Hugging Face (optional) |
- ๐ฅ๏ธ GPU: ๐ง 16 GiB (A4000). Peak ~12 GiB with
LOW_GPU_MEM (auto below 30 GiB: pre-quantized 4-bit 14B base); ~30 GiB without.
- โฑ๏ธ Runtime (ex. downloads/API): GPU ~9.4 min (+ API calls) ยท CPU N/A
- โ๏ธ Colab? โ Untested: loading the 4-bit 14B model in 12.7 GB RAM
- ๐พ Downloads: ๐ง ~30 GB (~11 GB with
LOW_GPU_MEM): Qwen/Qwen2.5-14B-Instruct 30 GB, ModelOrganismsForEM/Qwen2.5-14B-Instruct_bad-medical-advice 0.55 GB (LOW_GPU_MEM or Colab: unsloth/Qwen2.5-14B-Instruct-bnb-4bit 10 GB instead of the 30 GB)
- ๐ API keys:
- ๐ฅ Mandatory: OpenRouter (
OPENROUTER_API_KEY), for the autorater
- ๐ฉ Optional: Hugging Face (
HF_TOKEN), for higher rate limits and faster downloads from the Hub
- ๐ RAM: ~6.7 GB at peak
|
| 4.2Science of Misalignment |
โฌ |
โ |
<1 min |
๐ข |
โฌ ~100 MB |
<1 GB |
OpenRouter |
- ๐ฅ๏ธ GPU: โฌ Not needed.
- โฑ๏ธ Runtime (ex. downloads/API): CPU <1 min
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: โฌ ~100 MB
- ๐ API keys:
- ๐ฅ Mandatory: OpenRouter (
OPENROUTER_API_KEY), for the LLM API calls throughout the day (about $10 for a full run)
- ๐ RAM: <1 GB at peak
|
| 4.3Interpreting Reasoning Models |
๐ง 14 GiB |
~3.8 min (+ API calls) |
N/A |
โ |
๐จ ~5.4 GB |
~8 GB |
OpenRouter |
- ๐ฅ๏ธ GPU: ๐ง 16 GiB (A4000), peak ~14 GiB. Tight on 16 GiB.
- โฑ๏ธ Runtime (ex. downloads/API): GPU ~3.8 min (+ API calls) ยท CPU N/A
- โ๏ธ Colab? โ Untested: GPU tight (14.4 of 15 GiB)
- ๐พ Downloads: ๐จ ~5.4 GB: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B 3.6 GB, uzaymacar/blackmail-rollouts 1.2 GB, uzaymacar/math-rollouts 0.57 GB
- ๐ API keys:
- ๐ฅ Mandatory: OpenRouter (
OPENROUTER_API_KEY), for the API sections
- ๐ RAM: ~8 GB at peak
|
| 4.4LLM Psychology & Persona Vectors |
๐ง 29 GiB (11 GiB)* |
~38 min (+ API calls) |
N/A |
โ |
๐ฅ ~135 GB (~34 GB with LOW_GPU_MEM) |
~8.7 GB |
OpenRouter, Hugging Face |
- ๐ฅ๏ธ GPU: ๐ง 16 GiB (A4000). Peak ~11 GiB with
LOW_GPU_MEM (auto below 20 GiB: Gemma 2 9B in 8-bit, Qwen2.5-7B in 4-bit); ~29 GiB without.
- ๐ฅ Big GPU, Activation capping (ยง2): 24 GB+; errors on 16 GiB, skip it
- โฑ๏ธ Runtime (ex. downloads/API): GPU ~38 min (+ API calls) ยท CPU N/A
- โ๏ธ Colab? โ Untested: RAM ~9 GB not checked on Colab; activation capping needs 24 GB
- ๐พ Downloads: ๐ฅ ~135 GB (~34 GB with
LOW_GPU_MEM): Gemma 2 27B 55 GB, Qwen 3 32B 66 GB, Qwen2.5-7B 15 GB (LOW_GPU_MEM: Gemma 2 9B 19 GB + Qwen2.5-7B)
- ๐ API keys:
- ๐ฅ Mandatory: OpenRouter (
OPENROUTER_API_KEY), for generating responses and the autorater
- ๐ฅ Mandatory: Hugging Face (
HF_TOKEN), for Gemma 2 (request access: google/gemma-2-27b-it, google/gemma-2-9b-it (small GPUs))
- ๐ RAM: ~8.7 GB at peak
|
| 4.5Investigator Agents |
โฌ |
โ |
<1 min |
๐ข |
โฌ ~20 MB |
<1 GB |
OpenRouter |
- ๐ฅ๏ธ GPU: โฌ Not needed.
- โฑ๏ธ Runtime (ex. downloads/API): CPU <1 min
- โ๏ธ Colab? ๐ข Yes
- ๐พ Downloads: โฌ ~20 MB
- ๐ API keys:
- ๐ฅ Mandatory: OpenRouter (
OPENROUTER_API_KEY), for the LLM API calls throughout the day (about $25 for a full run)
- ๐ RAM: <1 GB at peak
|