minterp

tests/test_attribute.py

53 lines · View on GitHub · Linked from 6. Feature circuit

1import torch
2
3from minterp import attribute, tasks
4from minterp import sae as S
5from minterp.tap import Tap
6
7
8def setup(model, tok, people):
9    torch.manual_seed(0)
10    batch = tasks.batches(tok, people=people)[0]
11    sae = S.SAE(model.cfg.dim, 4 * model.cfg.dim, k=8).eval().requires_grad_(False)
12    measure = lambda logits, tap: tasks.logit_diff(logits, batch.right, batch.wrong)
13    return batch, sae, measure
14
15
16def test_the_dictionary_in_the_pass_changes_nothing(model, tok, people):
17    batch, sae, measure = setup(model, tok, people)
18    plain, _ = model(batch.clean)
19    z, err, gz, gerr, value = attribute.through(model, sae, batch.clean, 2, measure)
20    assert torch.allclose(value, tasks.logit_diff(plain, batch.right, batch.wrong), atol=1e-4)
21    assert z.shape == (len(batch), batch.clean.shape[1], sae.width) and gz.shape == z.shape and gerr.shape == err.shape
22    with Tap(model, edit={("resid", 2): attribute.rebuilt_with(sae, z, err)}, keep=()):
23        again, _ = model(batch.clean)
24    assert torch.allclose(again, plain, atol=1e-4)
25    assert not any(p.requires_grad for p in model.parameters())  # the model is still frozen
26
27
28def test_the_gradient_predicts_a_small_change_and_not_exactly_a_large_one(model, tok, people):
29    batch, sae, measure = setup(model, tok, people)
30    z, err, gz, gerr, value = attribute.through(model, sae, batch.clean, 2, measure)
31    torch.manual_seed(1)
32    with Tap(model, keep=("resid",)) as tap:
33        model(batch.clean)
34    step = torch.randn_like(err) * tap["resid", 2].std()  # a direction of the stream's own size at that depth
35
36    def moved(size):
37        with torch.no_grad(), Tap(model, edit={("resid", 2): attribute.rebuilt_with(sae, z, err + size * step)}, keep=()):
38            logits, _ = model(batch.clean)
39        return tasks.logit_diff(logits, batch.right, batch.wrong) - value
40
41    predicted = (step * gerr).sum((1, 2))
42    small, large = moved(1e-3) / 1e-3, moved(1.0)
43    off_small, off_large = (small - predicted).abs().mean(), (large - predicted).abs().mean()
44    assert off_small < 0.1 * predicted.abs().mean()  # first order: right for a small step
45    assert off_large > 5 * off_small  # and off for a large one
46
47
48def test_a_position_cannot_matter_to_an_earlier_one(model, tok, people):
49    batch, sae, _ = setup(model, tok, people)
50    early = lambda logits, tap: logits[:, 3, 0]  # a number read at position 3
51    _, _, gz, gerr, _ = attribute.through(model, sae, batch.clean, 2, early)
52    assert gerr[:, 4:].abs().max() == 0 and gz[:, 4:].abs().max() == 0
53    assert gerr[:, :4].abs().max() > 0