Skip to content

Xproportion official

auto_circuit.metrics.official_circuits.circuits.tracr.xproportion_official

Classes

Functions

tracr_xproportion_acdc_edges 

tracr_xproportion_acdc_edges(model: PatchableModel, token_positions: bool = False, word_idxs: Dict[str, int] = {}, seq_start_idx: int = 0) -> Set[Edge]

The canonical circuit for tracr-reverse according to Conmy et al. (2023). Based on the ACDC repo.

As discussed in Miller et al. (forthcoming), this circuit is (intended to be) the set of edges that must be preserved when Zero Ablation is used.

Parameters:

Name Type Description Default
model PatchableModel

A patchable TransformerLens tracr-reverse HookedTransformer model.

 required
token_positions bool

Whether to distinguish between token positions when returning the set of circuit edges. If True, we require that the model has seq_len not None (ie. separate edges for each token position) and that word_idxs is provided.

 False
word_idxs Dict[str, int]

A dictionary defining the index of specific named tokens in the circuit definition. This variable is not used in this circuit, instead we assume a sequence of length 6 (including BOS).

 {}
seq_start_idx int

Offset to add to all of the token positions in word_idxs. This is useful when using KV caching to skip the common prefix of the prompt.

 0

Returns:

Type Description
Set[Edge]

The set of edges in the circuit.
Note

The sequence positions assume prompts of length 6 (including BOS), as in tracr/tracr_reverse_len_5_prompts.json

Source code in auto_circuit/metrics/official_circuits/circuits/tracr/xproportion_official.py 
 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
def tracr_xproportion_acdc_edges(
    model: PatchableModel,
    token_positions: bool = False,
    word_idxs: Dict[str, int] = {},
    seq_start_idx: int = 0,
) -> Set[Edge]:
    """
    The canonical circuit for tracr-reverse according to
    [Conmy et al. (2023)](https://arxiv.org/abs/2304.14997). Based on the
    [ACDC repo](https://github.com/ArthurConmy/Automatic-Circuit-Discovery/blob/main/acdc/tracr_task/utils.py).

    As discussed in Miller et al. (forthcoming), this circuit is (intended to be) the
    set of edges that must be preserved when Zero Ablation is used.

    Args:
        model: A patchable TransformerLens tracr-reverse `HookedTransformer` model.
        token_positions: Whether to distinguish between token positions when returning
            the set of circuit edges. If `True`, we require that the `model` has
            `seq_len` not `None` (ie. separate edges for each token position) and that
            `word_idxs` is provided.
        word_idxs: A dictionary defining the index of specific named tokens in the
            circuit definition. This variable is not used in this circuit, instead we
            assume a sequence of length 6 (including BOS).
        seq_start_idx: Offset to add to all of the token positions in `word_idxs`.
            This is useful when using KV caching to skip the common prefix of the
            prompt.

    Returns:
        The set of edges in the circuit.

    Note:
        The sequence positions assume prompts of length 6 (including BOS), as in
        tracr/tracr_reverse_len_5_prompts.json
    """
    assert model.cfg.model_name == "tracr-xproportion"
    assert model.separate_qkv

    tok_seq_pos = [1, 2, 3, 4, 5]
    edges_present: Dict[str, List[int]] = {}

    edges_present["A1.0->Resid End"] = tok_seq_pos
    edges_present["Resid Start->A1.0.Q"] = tok_seq_pos
    edges_present["Resid Start->A1.0.K"] = tok_seq_pos
    edges_present["MLP 0->A1.0.V"] = tok_seq_pos
    edges_present["Resid Start->MLP 0"] = tok_seq_pos

    true_edges: Set[Edge] = set()
    for edge in model.edges:
        if edge.name in edges_present.keys():
            if token_positions:
                assert edge.seq_idx is not None
                if (edge.seq_idx + seq_start_idx) in edges_present[edge.name]:
                    true_edges.add(edge)
            else:
                true_edges.add(edge)
    return true_edges

tracr_xproportion_official_edges 

tracr_xproportion_official_edges(model: PatchableModel, token_positions: bool = False, word_idxs: Dict[str, int] = {}, seq_start_idx: int = 0) -> Set[Edge]

The canonical circuit for tracr-reverse according to Miller et al. (Forthcoming). As discussed in the paper, this circuit is the set of edges that must be preserved when Resample Ablation is used.

Parameters:

Name Type Description Default
model PatchableModel

A patchable TransformerLens tracr-reverse HookedTransformer model.

 required
token_positions bool

Whether to distinguish between token positions when returning the set of circuit edges. If True, we require that the model has seq_len not None (ie. separate edges for each token position) and that word_idxs is provided.

 False
word_idxs Dict[str, int]

A dictionary defining the index of specific named tokens in the circuit definition. This variable is not used in this circuit, instead we assume a sequence of length 6 (including BOS).

 {}
seq_start_idx int

Offset to add to all of the token positions in word_idxs. This is useful when using KV caching to skip the common prefix of the prompt.

 0

Returns:

Type Description
Set[Edge]

The set of edges in the circuit.
Note

The sequence positions assume prompts of length 6 (including BOS), as in tracr/tracr_xproportion_len_5_prompts.json

Source code in auto_circuit/metrics/official_circuits/circuits/tracr/xproportion_official.py 
 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
def tracr_xproportion_official_edges(
    model: PatchableModel,
    token_positions: bool = False,
    word_idxs: Dict[str, int] = {},
    seq_start_idx: int = 0,
) -> Set[Edge]:
    """
    The canonical circuit for tracr-reverse according to Miller et al. (Forthcoming).
    As discussed in the paper, this circuit is the set of edges that must be preserved
    when Resample Ablation is used.

    Args:
        model: A patchable TransformerLens tracr-reverse `HookedTransformer` model.
        token_positions: Whether to distinguish between token positions when returning
            the set of circuit edges. If `True`, we require that the `model` has
            `seq_len` not `None` (ie. separate edges for each token position) and that
            `word_idxs` is provided.
        word_idxs: A dictionary defining the index of specific named tokens in the
            circuit definition. This variable is not used in this circuit, instead we
            assume a sequence of length 6 (including BOS).
        seq_start_idx: Offset to add to all of the token positions in `word_idxs`.
            This is useful when using KV caching to skip the common prefix of the
            prompt.

    Returns:
        The set of edges in the circuit.

    Note:
        The sequence positions assume prompts of length 6 (including BOS), as in
        tracr/tracr_xproportion_len_5_prompts.json
    """
    assert model.cfg.model_name == "tracr-xproportion"
    assert model.separate_qkv

    tok_seq_pos = [1, 2, 3, 4, 5]
    edges_present: Dict[str, List[int]] = {}
    edges_present["Resid Start->MLP 0"] = tok_seq_pos
    edges_present["MLP 0->A1.0.V"] = tok_seq_pos
    edges_present["A1.0->Resid End"] = tok_seq_pos

    true_edges: Set[Edge] = set()
    for edge in model.edges:
        if edge.name in edges_present.keys():
            if token_positions:
                assert edge.seq_idx is not None
                if (edge.seq_idx + seq_start_idx) in edges_present[edge.name]:
                    true_edges.add(edge)
            else:
                true_edges.add(edge)
    return true_edges