README.md

---
tags:
  - sparse-autoencoder
  - crosscoder
  - interpretability
  - qwen2
  - mechanistic-interpretability
  - dictionary-learning
license: mit
---

# dfc-D8k-excl3-freeexcl-k90

A **DFC** sparse crosscoder trained to compare layer-13 activations between:
- **Model A (ToolRL)**: `chengq9/ToolRL-Qwen2.5-3B` — fine-tuned with tool-use reinforcement learning
- **Model B (Base)**: `Qwen/Qwen2.5-3B` — vanilla base model

## What is this?

This model learns a sparse dictionary of features from the internal representations of two language models. By comparing which features activate for which model, we can identify:
- **What the ToolRL fine-tuning changed** (A-exclusive features)
- **What remained the same** (shared features)
- **What the base model does that ToolRL suppressed** (B-exclusive features)

## Model Architecture

Dedicated Feature CrossCoder with partitioned dictionary (0.03/0.03 A/B exclusive)

| Parameter | Value |
|-----------|-------|
| Dictionary size | 8192 |
| Top-k active features | 90 |
| Layer | 13 (middle layer of Qwen2.5-3B) |
| Activation dimension | 2048 |
| A-exclusive features | 245 (0.03) |
| B-exclusive features | 245 (0.03) |
| Shared features | 7702 |

### How it works

1. **Encode**: Takes stacked activations `(batch, 2, 2048)` from both models, applies per-model encoder weights, sums across models, and selects the top-90 features via ReLU + top-k.
2. **Decode**: Reconstructs per-model activations from the sparse feature vector using per-model decoder weights.
3. **Partition masks** (DFC only): Hard binary masks zero out encoder/decoder weights to enforce that exclusive features cannot be used by the wrong model.

## Training

| Parameter | Value |
|-----------|-------|
| Loss function | MSE + L1 on shared features only (coef: 1e-3); exclusive features unpenalized |
| Training steps | 9000 |
| Learning rate | 1e-4 |
| Batch size | 1024 |
| Sparsity coefficient (shared) | 1e-3 |
| Exclusive sparsity coefficient | 0 |
| Optimizer | Adam (grad clip 1.0) |
| W&B project | `dfc-crosscoder-sweep` |

### Training Data

- **FineWeb**: ~40,000 general web text samples (from `HuggingFaceFW/fineweb` sample-10BT)
- **ToolRL**: ~40,000 tool-use conversation samples (from `emrecanacikgoz/ToolRL`, cycled)
- Activations extracted from layer 13, last token per sample
- Both datasets concatenated and z-score normalized

## Usage

### Quick Start

```python
import torch
from huggingface_hub import hf_hub_download

# Download model files
repo_id = "antebe1/dfc-D8k-excl3-freeexcl-k90"
for fname in ["model.pt", "config.json", "dfc.py"]:
    hf_hub_download(repo_id=repo_id, filename=fname, local_dir="./model")

# Load the crosscoder
import sys; sys.path.insert(0, "./model")
from dfc import DFCCrossCoder

dfc = DFCCrossCoder.load("./model", device="cuda")
print(f"Loaded: dict_size={dfc.dict_size}, k={dfc.k}")
```

### Extract Features from Real Models

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load both models
model_a = AutoModelForCausalLM.from_pretrained("chengq9/ToolRL-Qwen2.5-3B", device_map="cuda:0")
model_b = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-3B", device_map="cuda:1")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-3B")

# Get activations from layer 13
# NOTE: hidden_states[0] = embeddings, hidden_states[i] = output of layer i-1
#       so layer 13 activations are at index 13+1
text = "Use the search tool to find recent papers on RLHF"
inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    out_a = model_a(**inputs.to("cuda:0"), output_hidden_states=True)
    out_b = model_b(**inputs.to("cuda:1"), output_hidden_states=True)
    act_a = out_a.hidden_states[13 + 1][:, -1, :]  # last token, layer 13
    act_b = out_b.hidden_states[13 + 1][:, -1, :]

# Stack and encode
activations = torch.stack([act_a.cpu(), act_b.cpu()], dim=1)  # (1, 2, 2048)
features = dfc.encode(activations.to(dfc.W_enc.device))

print(f"Active features: {(features > 0).sum().item()} / {dfc.dict_size}")
```

### Analyze Partitions (DFC only)

```python
stats = dfc.feature_stats(features)
print(f"L0 total:    {stats['l0_total']:.1f}")
print(f"L0 A-excl:   {stats['l0_a_excl']:.1f}")
print(f"L0 B-excl:   {stats['l0_b_excl']:.1f}")
print(f"L0 shared:   {stats['l0_shared']:.1f}")

# Check reconstruction quality
recon, feats = dfc(activations.to(dfc.W_enc.device))
mse = torch.nn.functional.mse_loss(recon.cpu(), activations)
print(f"Reconstruction MSE: {mse.item():.6f}")
```

## Files

| File | Description |
|------|-------------|
| `model.pt` | PyTorch state dict (encoder/decoder weights + partition masks) |
| `config.json` | Architecture config: dict_size, k, partition sizes (n_a, n_b) |
| `hparams.json` | Full training hyperparameters including loss, lr, steps, etc. |
| `dfc.py` | `DFCCrossCoder` class definition — required to load model.pt |
| `demo.py` | Feature extraction demo (works with downloaded model) |
| `requirements.txt` | Python dependencies |

## Part of a Sweep

This model is one of 48 models in a hyperparameter sweep. See the full collection:

| Axis | Values |
|------|--------|
| k (top-k) | 45, 90, 160 |
| dict_size | 8,192 / 16,384 |
| Architecture | DFC (partitioned) / CrossCoder (all shared) |
| Exclusive % (DFC) | 3%, 5%, 10% |
| Exclusive sparsity | 1e-3 (penalized) / 0 (free) |
| CrossCoder L1 | with / without |

## Citation

```bibtex
@misc{dfc-D8k-excl3-freeexcl-k90,
  title={DFC CrossCoder: ToolRL vs Base Qwen2.5-3B},
  author={Andre Shportko},
  year={2026},
  url={https://huggingface.co/antebe1/dfc-D8k-excl3-freeexcl-k90}
}
```