R1-Interp

Interpretability toolkit for DeepSeek-R1-Distill-Qwen-1.5B — the first complete circuit-analysis pipeline for a reasoning-distilled language model.

R1-Interp trains skip transcoders on all 28 MLP layers of R1-Distill-Qwen-1.5B, generates Anthropic-style attribution graphs, and runs network motif analysis to compare circuit structure between reasoning and non-reasoning models.

Why This Project

Reasoning-distilled models like DeepSeek-R1 represent a new paradigm in LLM capabilities, but we don't yet understand how reasoning distillation changes a model's internal circuitry. R1-Interp aims to answer:

• Does reasoning distillation change circuit structure, or just feature usage?
• Are there distinct motif signatures in <think> reasoning traces vs. final answers?
• How do R1-distill circuits compare to base Qwen and Claude Haiku?

Pipeline Overview

Phase	Script	Description
0	setup_env.py, prepare_data.py	Environment setup, data download & tokenization
1	sweep.py	Hyperparameter sweep on 3 pilot layers (2, 14, 26)
2	train.py	Full transcoder training across all 28 layers
3	collect_activations.py, autointerp.py	Feature activation collection & Claude API labeling
4	generate_graphs.py, motif_analysis.py, compare_models.py	Attribution graphs, motif census, cross-model comparison

Quick Start

# Install
pip install -e ".[dev]"

# Verify environment
python scripts/setup_env.py

# Run tests
pytest tests/

Project Structure

r1_interp/                  # Core package
├── config.py               # Frozen dataclasses for all pipeline configs
├── prompts.py              # 24 benchmark prompts across 7 categories
├── model_registry.py       # R1-distill & base Qwen specs for circuit-motifs
└── utils.py                # Device detection, path helpers, timing

configs/                    # YAML training configs
├── sweep_pilot.yaml        # Phase 1: 3×3 sweep grid (expansion × topk)
├── train_full.yaml         # Phase 2: all 28 layers (update after sweep)
├── autointerp.yaml         # Phase 3: Claude API feature labeling
└── graph_generation.yaml   # Phase 4: attribution graph parameters

scripts/                    # One script per pipeline stage
├── setup_env.py            # Check deps, clone sparsify, download model
├── prepare_data.py         # Pre-tokenize OpenR1-Math corpus
├── train.py                # Thin wrapper around sparsify (CLI + Python API)
├── sweep.py                # Grid search with Pareto-optimal selection
├── evaluate.py             # CE loss increase, dead features, L0
├── collect_activations.py  # Top-N activating examples per feature
├── autointerp.py           # Rate-limited Claude API labeling with resume
├── generate_graphs.py      # circuit-tracer attribution + Neuronpedia export
├── motif_analysis.py       # circuit-motifs pipeline wrapper
└── compare_models.py       # R1 vs Qwen + think-vs-answer comparison

data/prompts/
└── benchmark_prompts.json  # 24 prompts: arithmetic, multihop, reasoning,
                            # factual_recall, safety, reasoning_failure, baseline

tests/                      # 47 tests
├── test_config.py          # Config properties, sweep generation, YAML roundtrip
├── test_prompts.py         # Unique IDs, category coverage, JSON roundtrip
├── test_model_registry.py  # Model specs, circuit-motifs registration
└── test_graph_compat.py    # Neuronpedia JSON ↔ circuit-motifs compatibility

Key Dependencies

Dependency	Purpose
EleutherAI/sparsify	Transcoder training framework
circuit-tracer	Attribution graph generation
circuit-motifs	Network motif analysis for circuit comparison
wandb	Experiment tracking

Training Configuration

Parameter	Default	Notes
Expansion factor	32×, 64×, 128×	Sweep candidates (of hidden dim 1536)
TopK	32, 64, 128	Sparsity parameter
Skip connection	Yes	Dunefsky et al. skip transcoders
Training tokens	200M	Per layer
Faithfulness target	<0.1 nats CE increase	Ideally <0.05
Dead features target	<10%

Benchmark Prompts

24 prompts across 7 categories designed for circuit comparison:

• arithmetic (5) — addition, multiplication, algebra, order of operations
• multihop (4) — Dallas→Texas→Austin style multi-step retrieval
• reasoning (4) — snail problem, syllogism, missing dollar, bat-and-ball
• factual_recall (4) — simple facts for baseline comparison
• safety (3) — refusal triggers and boundary cases
• reasoning_failure (2) — problems the model is likely to get wrong
• baseline (2) — text completion (no reasoning required)

Planned Comparisons

1. R1-distill vs. base Qwen-1.5B — Does reasoning distillation change circuit wiring?
2. <think> phase vs. answer phase — Are reasoning traces structurally different?
3. R1-distill vs. Claude Haiku — Cross-architecture motif comparison
4. Correct vs. incorrect reasoning — Circuit signatures of failure

License

MIT