File size: 4,717 Bytes
88b5d86
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
# K-FAC Memorization Suppression

Reproduction of ["From Memorization to Reasoning in the Spectrum of Loss Curvature"](https://github.com/goodfire-ai/memorization_kfac) with extended experiments on modified importance formulas.

## Overview

This project implements K-FAC (Kronecker-Factored Approximate Curvature) based weight editing to suppress verbatim memorization in language models while preserving general capabilities.

**Key insight:** The Fisher Information Matrix, approximated by K-FAC, reveals directions in weight space associated with memorization (low curvature) vs. generalization (high curvature). By removing low-curvature components, we can suppress memorization.

## Project Goal

1. **Reproduce** the paper's K-FAC method on OLMo-2 1B
2. **Compare** the original importance formula with a modified version:
   - **Original:** $\Pi_{ij} = \lambda_i \cdot \mu_j$
   - **Modified:** $\Pi_{ij} = \lambda_i \cdot \mu_j \cdot |C_{ij}|^2$

## Installation

```bash
pip install -r requirements.txt
```

## Project Structure

```
├── src/
│   ├── kfac_collector.py    # Collect K-FAC statistics (A, G matrices)
│   ├── kfac_editor.py       # Weight editing via eigendecomposition
│   ├── evaluate.py          # Memorization and perplexity metrics
│   └── mine_memorized.py    # Mine memorized sequences from training data
├── notebooks/
│   ├── 01_collect_kfac.ipynb      # Colab: K-FAC collection (~2h on A100)
│   ├── 02_mine_memorized.ipynb    # Colab: Find memorized sequences (~1h)
│   └── 03_experiments.ipynb       # Colab: Run experiments (~2h)
├── plans/
│   └── implementation_plan.md     # Detailed implementation plan
├── context/
│   ├── original_paper/            # Paper sections in markdown
│   └── REPRODUCTION_PLAN.md       # Initial reproduction plan
└── requirements.txt
```

## Quick Start

### Local Development

```python
from src.kfac_collector import KFACCollector, KFACConfig
from src.kfac_editor import KFACEditor, EditConfig
from src.evaluate import memorization_score, perplexity

# Load model
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("allenai/OLMo-2-1124-7B")
tokenizer = AutoTokenizer.from_pretrained("allenai/OLMo-2-1124-7B")

# Load pre-collected K-FAC stats
collector = KFACCollector.load("kfac_statistics.pt", model)
kfac_stats = collector.get_statistics()

# Apply K-FAC editing
config = EditConfig(energy_threshold=0.6, formula="original")
editor = KFACEditor(model, kfac_stats, config)
editor.edit_model()

# Evaluate
result = memorization_score(model, tokenizer, prefixes, suffixes)
print(f"Strict accuracy: {result.strict_accuracy*100:.1f}%")
```

### Running on Colab

1. **01_collect_kfac.ipynb** - Collect K-FAC statistics (~20M tokens, ~2h on A100)
2. **02_mine_memorized.ipynb** - Find memorized sequences from training data
3. **03_experiments.ipynb** - Run experiments and compare formulas

## Method

### K-FAC Statistics Collection

For each target MLP layer, we collect:
- **A**: Activation covariance matrix (input side)
- **G**: Gradient covariance matrix (output side)

These approximate the Fisher Information Matrix: $F_W \approx G \otimes A$

### Weight Editing

1. **Eigendecompose** A and G matrices
2. **Transform** weights to curvature basis: $C = U_G^T W U_A$
3. **Compute importance** using either formula
4. **Select** top components by cumulative energy (e.g., 60%)
5. **Reconstruct** edited weights: $W_{edited} = U_G (C \odot M) U_A^T$

### Importance Formulas

| Formula | Definition | Intuition |
|---------|------------|-----------|
| Original | $\Pi_{ij} = \lambda_i \cdot \mu_j$ | Pure curvature |
| Modified | $\Pi_{ij} = \lambda_i \cdot \mu_j \cdot |C_{ij}|^2$ | Curvature weighted by actual weight magnitude |

## Hyperparameters

| Parameter | 7B Model | 1B Model (estimated) |
|-----------|----------|---------------------|
| Target layers | 23, 24, 25 | 11, 12, 13 |
| Projections | gate_proj, up_proj | gate_proj, up_proj |
| Energy threshold | 60% | 60% |
| K-FAC tokens | ~20M | ~20M |

## Expected Results

Based on the paper (OLMo-2 1B):

| Metric | Baseline | After K-FAC |
|--------|----------|-------------|
| Dolma strict accuracy | ~98% | ~3% |
| Perplexity (Pile-10k) | ~23 | ~27 |

## References

- Paper: [From Memorization to Reasoning in the Spectrum of Loss Curvature](https://github.com/goodfire-ai/memorization_kfac)
- Original code: https://github.com/goodfire-ai/memorization_kfac
- Model: [OLMo-2](https://huggingface.co/allenai/OLMo-2-1124-7B)
- K-FAC: [Martens & Grosse, 2015](https://arxiv.org/abs/1503.05671)

## License

MIT