sae-lens 6.25.0__tar.gz → 6.26.1__tar.gz

{sae_lens-6.25.0 → sae_lens-6.26.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sae-lens
-Version: 6.25.0
+Version: 6.26.1
 Summary: Training and Analyzing Sparse Autoencoders (SAEs)
 License: MIT
 License-File: LICENSE

{sae_lens-6.25.0 → sae_lens-6.26.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "sae-lens"
-version = "6.25.0"
+version = "6.26.1"
 description = "Training and Analyzing Sparse Autoencoders (SAEs)"
 authors = ["Joseph Bloom"]
 readme = "README.md"
@@ -51,6 +51,7 @@ mkdocs-redirects = "^1.2.1"
 mkdocs-section-index = "^0.3.9"
 mkdocstrings = "^0.25.2"
 mkdocstrings-python = "^1.10.9"
+beautifulsoup4 = "^4.12.0"
 tabulate = "^0.9.0"
 ruff = "^0.7.4"
 eai-sparsify = "^1.1.1"

{sae_lens-6.25.0 → sae_lens-6.26.1}/sae_lens/__init__.py

@@ -1,5 +1,5 @@
 # ruff: noqa: E402
-__version__ = "6.25.0"
+__version__ = "6.26.1"
 
 import logging
 
@@ -21,6 +21,10 @@ from sae_lens.saes import (
     JumpReLUTrainingSAEConfig,
     JumpReLUTranscoder,
     JumpReLUTranscoderConfig,
+    MatchingPursuitSAE,
+    MatchingPursuitSAEConfig,
+    MatchingPursuitTrainingSAE,
+    MatchingPursuitTrainingSAEConfig,
     MatryoshkaBatchTopKTrainingSAE,
     MatryoshkaBatchTopKTrainingSAEConfig,
     SAEConfig,
@@ -113,6 +117,10 @@ __all__ = [
     "MatryoshkaBatchTopKTrainingSAEConfig",
     "TemporalSAE",
     "TemporalSAEConfig",
+    "MatchingPursuitSAE",
+    "MatchingPursuitTrainingSAE",
+    "MatchingPursuitSAEConfig",
+    "MatchingPursuitTrainingSAEConfig",
 ]
 
 
@@ -139,3 +147,7 @@ register_sae_class(
     "jumprelu_skip_transcoder", JumpReLUSkipTranscoder, JumpReLUSkipTranscoderConfig
 )
 register_sae_class("temporal", TemporalSAE, TemporalSAEConfig)
+register_sae_class("matching_pursuit", MatchingPursuitSAE, MatchingPursuitSAEConfig)
+register_sae_training_class(
+    "matching_pursuit", MatchingPursuitTrainingSAE, MatchingPursuitTrainingSAEConfig
+)

{sae_lens-6.25.0 → sae_lens-6.26.1}/sae_lens/cache_activations_runner.py

@@ -14,9 +14,9 @@ from transformer_lens.HookedTransformer import HookedRootModule
 
 from sae_lens import logger
 from sae_lens.config import CacheActivationsRunnerConfig
-from sae_lens.constants import DTYPE_MAP
 from sae_lens.load_model import load_model
 from sae_lens.training.activations_store import ActivationsStore
+from sae_lens.util import str_to_dtype
 
 
 def _mk_activations_store(
@@ -97,7 +97,7 @@ class CacheActivationsRunner:
         bytes_per_token = (
             self.cfg.d_in * self.cfg.dtype.itemsize
             if isinstance(self.cfg.dtype, torch.dtype)
-            else DTYPE_MAP[self.cfg.dtype].itemsize
+            else str_to_dtype(self.cfg.dtype).itemsize
         )
         total_training_tokens = self.cfg.n_seq_in_dataset * self.context_size
         total_disk_space_gb = total_training_tokens * bytes_per_token / 10**9

{sae_lens-6.25.0 → sae_lens-6.26.1}/sae_lens/config.py

@@ -17,9 +17,14 @@ from datasets import (
 )
 
 from sae_lens import __version__, logger
-from sae_lens.constants import DTYPE_MAP
+
+# keeping this unused import since some SAELens deps import DTYPE_MAP from config
+from sae_lens.constants import (
+    DTYPE_MAP,  # noqa: F401  # pyright: ignore[reportUnusedImport]
+)
 from sae_lens.registry import get_sae_training_class
 from sae_lens.saes.sae import TrainingSAEConfig
+from sae_lens.util import str_to_dtype
 
 if TYPE_CHECKING:
     pass
@@ -563,7 +568,7 @@ class CacheActivationsRunnerConfig:
 
     @property
     def bytes_per_token(self) -> int:
-        return self.d_in * DTYPE_MAP[self.dtype].itemsize
+        return self.d_in * str_to_dtype(self.dtype).itemsize
 
     @property
     def n_tokens_in_buffer(self) -> int:

{sae_lens-6.25.0 → sae_lens-6.26.1}/sae_lens/constants.py

@@ -11,6 +11,14 @@ DTYPE_MAP = {
     "torch.bfloat16": torch.bfloat16,
 }
 
+# Reverse mapping from torch.dtype to canonical string format
+DTYPE_TO_STR = {
+    torch.float32: "float32",
+    torch.float64: "float64",
+    torch.float16: "float16",
+    torch.bfloat16: "bfloat16",
+}
+
 
 SPARSITY_FILENAME = "sparsity.safetensors"
 SAE_WEIGHTS_FILENAME = "sae_weights.safetensors"

{sae_lens-6.25.0 → sae_lens-6.26.1}/sae_lens/loading/pretrained_sae_loaders.py

@@ -12,11 +12,9 @@ from huggingface_hub import hf_hub_download, hf_hub_url
 from huggingface_hub.utils import EntryNotFoundError, build_hf_headers
 from packaging.version import Version
 from safetensors import safe_open
-from safetensors.torch import load_file
 
 from sae_lens import logger
 from sae_lens.constants import (
-    DTYPE_MAP,
     SAE_CFG_FILENAME,
     SAE_WEIGHTS_FILENAME,
     SPARSIFY_WEIGHTS_FILENAME,
@@ -28,7 +26,7 @@ from sae_lens.loading.pretrained_saes_directory import (
     get_repo_id_and_folder_name,
 )
 from sae_lens.registry import get_sae_class
-from sae_lens.util import filter_valid_dataclass_fields
+from sae_lens.util import filter_valid_dataclass_fields, str_to_dtype
 
 LLM_METADATA_KEYS = {
     "model_name",
@@ -51,6 +49,21 @@ LLM_METADATA_KEYS = {
 }
 
 
+def load_safetensors_weights(
+    path: str | Path, device: str = "cpu", dtype: torch.dtype | str | None = None
+) -> dict[str, torch.Tensor]:
+    """Load safetensors weights and optionally convert to a different dtype"""
+    loaded_weights = {}
+    dtype = str_to_dtype(dtype) if isinstance(dtype, str) else dtype
+    with safe_open(path, framework="pt", device=device) as f:
+        for k in f.keys():  # noqa: SIM118
+            weight = f.get_tensor(k)
+            if dtype is not None:
+                weight = weight.to(dtype=dtype)
+            loaded_weights[k] = weight
+    return loaded_weights
+
+
 # loaders take in a release, sae_id, device, and whether to force download, and returns a tuple of config, state_dict, and log sparsity
 class PretrainedSaeHuggingfaceLoader(Protocol):
     def __call__(
@@ -341,7 +354,7 @@ def read_sae_components_from_disk(
     Given a loaded dictionary and a path to a weight file, load the weights and return the state_dict.
     """
     if dtype is None:
-        dtype = DTYPE_MAP[cfg_dict["dtype"]]
+        dtype = str_to_dtype(cfg_dict["dtype"])
 
     state_dict = {}
     with safe_open(weight_path, framework="pt", device=device) as f:  # type: ignore
@@ -695,7 +708,9 @@ def gemma_3_sae_huggingface_loader(
         force_download=force_download,
     )
 
-    raw_state_dict = load_file(sae_path, device=device)
+    raw_state_dict = load_safetensors_weights(
+        sae_path, device=device, dtype=cfg_dict.get("dtype")
+    )
 
     with torch.no_grad():
         w_dec = raw_state_dict["w_dec"]
@@ -782,11 +797,13 @@ def get_goodfire_huggingface_loader(
     )
     raw_state_dict = torch.load(sae_path, map_location=device)
 
+    target_dtype = str_to_dtype(cfg_dict.get("dtype", "float32"))
+
     state_dict = {
-        "W_enc": raw_state_dict["encoder_linear.weight"].T,
-        "W_dec": raw_state_dict["decoder_linear.weight"].T,
-        "b_enc": raw_state_dict["encoder_linear.bias"],
-        "b_dec": raw_state_dict["decoder_linear.bias"],
+        "W_enc": raw_state_dict["encoder_linear.weight"].T.to(dtype=target_dtype),
+        "W_dec": raw_state_dict["decoder_linear.weight"].T.to(dtype=target_dtype),
+        "b_enc": raw_state_dict["encoder_linear.bias"].to(dtype=target_dtype),
+        "b_dec": raw_state_dict["decoder_linear.bias"].to(dtype=target_dtype),
     }
 
     return cfg_dict, state_dict, None
@@ -889,26 +906,19 @@ def llama_scope_sae_huggingface_loader(
         force_download=force_download,
     )
 
-    # Load the weights using load_file instead of safe_open
-    state_dict_loaded = load_file(sae_path, device=device)
+    state_dict_loaded = load_safetensors_weights(
+        sae_path, device=device, dtype=cfg_dict.get("dtype")
+    )
 
     # Convert and organize the weights
     state_dict = {
-        "W_enc": state_dict_loaded["encoder.weight"]
-        .to(dtype=DTYPE_MAP[cfg_dict["dtype"]])
-        .T,
-        "W_dec": state_dict_loaded["decoder.weight"]
-        .to(dtype=DTYPE_MAP[cfg_dict["dtype"]])
-        .T,
-        "b_enc": state_dict_loaded["encoder.bias"].to(
-            dtype=DTYPE_MAP[cfg_dict["dtype"]]
-        ),
-        "b_dec": state_dict_loaded["decoder.bias"].to(
-            dtype=DTYPE_MAP[cfg_dict["dtype"]]
-        ),
+        "W_enc": state_dict_loaded["encoder.weight"].T,
+        "W_dec": state_dict_loaded["decoder.weight"].T,
+        "b_enc": state_dict_loaded["encoder.bias"],
+        "b_dec": state_dict_loaded["decoder.bias"],
         "threshold": torch.ones(
             cfg_dict["d_sae"],
-            dtype=DTYPE_MAP[cfg_dict["dtype"]],
+            dtype=str_to_dtype(cfg_dict["dtype"]),
             device=cfg_dict["device"],
         )
         * cfg_dict["jump_relu_threshold"],
@@ -1219,26 +1229,17 @@ def llama_scope_r1_distill_sae_huggingface_loader(
         force_download=force_download,
     )
 
-    # Load the weights using load_file instead of safe_open
-    state_dict_loaded = load_file(sae_path, device=device)
+    state_dict_loaded = load_safetensors_weights(
+        sae_path, device=device, dtype=cfg_dict.get("dtype")
+    )
 
     # Convert and organize the weights
     state_dict = {
-        "W_enc": state_dict_loaded["encoder.weight"]
-        .to(dtype=DTYPE_MAP[cfg_dict["dtype"]])
-        .T,
-        "W_dec": state_dict_loaded["decoder.weight"]
-        .to(dtype=DTYPE_MAP[cfg_dict["dtype"]])
-        .T,
-        "b_enc": state_dict_loaded["encoder.bias"].to(
-            dtype=DTYPE_MAP[cfg_dict["dtype"]]
-        ),
-        "b_dec": state_dict_loaded["decoder.bias"].to(
-            dtype=DTYPE_MAP[cfg_dict["dtype"]]
-        ),
-        "threshold": state_dict_loaded["log_jumprelu_threshold"]
-        .to(dtype=DTYPE_MAP[cfg_dict["dtype"]])
-        .exp(),
+        "W_enc": state_dict_loaded["encoder.weight"].T,
+        "W_dec": state_dict_loaded["decoder.weight"].T,
+        "b_enc": state_dict_loaded["encoder.bias"],
+        "b_dec": state_dict_loaded["decoder.bias"],
+        "threshold": state_dict_loaded["log_jumprelu_threshold"].exp(),
     }
 
     # No sparsity tensor for Llama Scope SAEs
@@ -1358,34 +1359,34 @@ def sparsify_disk_loader(
     cfg_dict = get_sparsify_config_from_disk(path, device, cfg_overrides)
 
     weight_path = Path(path) / SPARSIFY_WEIGHTS_FILENAME
-    state_dict_loaded = load_file(weight_path, device=device)
-
-    dtype = DTYPE_MAP[cfg_dict["dtype"]]
+    state_dict_loaded = load_safetensors_weights(
+        weight_path, device=device, dtype=cfg_dict.get("dtype")
+    )
 
     W_enc = (
         state_dict_loaded["W_enc"]
         if "W_enc" in state_dict_loaded
         else state_dict_loaded["encoder.weight"].T
-    ).to(dtype)
+    )
 
     if "W_dec" in state_dict_loaded:
-        W_dec = state_dict_loaded["W_dec"].T.to(dtype)
+        W_dec = state_dict_loaded["W_dec"].T
     else:
-        W_dec = state_dict_loaded["decoder.weight"].T.to(dtype)
+        W_dec = state_dict_loaded["decoder.weight"].T
 
     if "b_enc" in state_dict_loaded:
-        b_enc = state_dict_loaded["b_enc"].to(dtype)
+        b_enc = state_dict_loaded["b_enc"]
     elif "encoder.bias" in state_dict_loaded:
-        b_enc = state_dict_loaded["encoder.bias"].to(dtype)
+        b_enc = state_dict_loaded["encoder.bias"]
     else:
-        b_enc = torch.zeros(cfg_dict["d_sae"], dtype=dtype, device=device)
+        b_enc = torch.zeros(cfg_dict["d_sae"], dtype=W_dec.dtype, device=device)
 
     if "b_dec" in state_dict_loaded:
-        b_dec = state_dict_loaded["b_dec"].to(dtype)
+        b_dec = state_dict_loaded["b_dec"]
     elif "decoder.bias" in state_dict_loaded:
-        b_dec = state_dict_loaded["decoder.bias"].to(dtype)
+        b_dec = state_dict_loaded["decoder.bias"]
     else:
-        b_dec = torch.zeros(cfg_dict["d_in"], dtype=dtype, device=device)
+        b_dec = torch.zeros(cfg_dict["d_in"], dtype=W_dec.dtype, device=device)
 
     state_dict = {"W_enc": W_enc, "b_enc": b_enc, "W_dec": W_dec, "b_dec": b_dec}
     return cfg_dict, state_dict
@@ -1616,7 +1617,9 @@ def mwhanna_transcoder_huggingface_loader(
     )
 
     # Load weights from safetensors
-    state_dict = load_file(file_path, device=device)
+    state_dict = load_safetensors_weights(
+        file_path, device=device, dtype=cfg_dict.get("dtype")
+    )
     state_dict["W_enc"] = state_dict["W_enc"].T
 
     return cfg_dict, state_dict, None
@@ -1700,8 +1703,12 @@ def mntss_clt_layer_huggingface_loader(
         force_download=force_download,
     )
 
-    encoder_state_dict = load_file(encoder_path, device=device)
-    decoder_state_dict = load_file(decoder_path, device=device)
+    encoder_state_dict = load_safetensors_weights(
+        encoder_path, device=device, dtype=cfg_dict.get("dtype")
+    )
+    decoder_state_dict = load_safetensors_weights(
+        decoder_path, device=device, dtype=cfg_dict.get("dtype")
+    )
 
     with torch.no_grad():
         state_dict = {
@@ -1844,7 +1851,9 @@ def temporal_sae_huggingface_loader(
     )
 
     # Load checkpoint from safetensors
-    state_dict_raw = load_file(ckpt_path, device=device)
+    state_dict_raw = load_safetensors_weights(
+        ckpt_path, device=device, dtype=cfg_dict.get("dtype")
+    )
 
     # Convert to SAELens naming convention
     # TemporalSAE uses: D (decoder), E (encoder), b (bias), attn_layers.*

{sae_lens-6.25.0 → sae_lens-6.26.1}/sae_lens/pretrained_saes.yaml

@@ -9072,150 +9072,150 @@ gemma-scope-2-27b-it-transcoders-all:
   - id: layer_5_width_262k_l0_small_affine
     path: transcoder_all/layer_5_width_262k_l0_small_affine
     l0: 12
-  # - id: layer_60_width_16k_l0_big
-  #   path: transcoder_all/layer_60_width_16k_l0_big
-  #   l0: 120
-  # - id: layer_60_width_16k_l0_big_affine
-  #   path: transcoder_all/layer_60_width_16k_l0_big_affine
-  #   l0: 120
-  # - id: layer_60_width_16k_l0_small
-  #   path: transcoder_all/layer_60_width_16k_l0_small
-  #   l0: 20
-  # - id: layer_60_width_16k_l0_small_affine
-  #   path: transcoder_all/layer_60_width_16k_l0_small_affine
-  #   l0: 20
-  # - id: layer_60_width_262k_l0_big
-  #   path: transcoder_all/layer_60_width_262k_l0_big
-  #   l0: 120
-  # - id: layer_60_width_262k_l0_big_affine
-  #   path: transcoder_all/layer_60_width_262k_l0_big_affine
-  #   l0: 120
-  # - id: layer_60_width_262k_l0_small
-  #   path: transcoder_all/layer_60_width_262k_l0_small
-  #   l0: 20
-  # - id: layer_60_width_262k_l0_small_affine
-  #   path: transcoder_all/layer_60_width_262k_l0_small_affine
-  #   l0: 20
-  # - id: layer_61_width_16k_l0_big
-  #   path: transcoder_all/layer_61_width_16k_l0_big
-  #   l0: 120
-  # - id: layer_61_width_16k_l0_big_affine
-  #   path: transcoder_all/layer_61_width_16k_l0_big_affine
-  #   l0: 120
-  # - id: layer_61_width_16k_l0_small
-  #   path: transcoder_all/layer_61_width_16k_l0_small
-  #   l0: 20
-  # - id: layer_61_width_16k_l0_small_affine
-  #   path: transcoder_all/layer_61_width_16k_l0_small_affine
-  #   l0: 20
-  # - id: layer_61_width_262k_l0_big
-  #   path: transcoder_all/layer_61_width_262k_l0_big
-  #   l0: 120
-  # - id: layer_61_width_262k_l0_big_affine
-  #   path: transcoder_all/layer_61_width_262k_l0_big_affine
-  #   l0: 120
-  # - id: layer_61_width_262k_l0_small
-  #   path: transcoder_all/layer_61_width_262k_l0_small
-  #   l0: 20
-  # - id: layer_61_width_262k_l0_small_affine
-  #   path: transcoder_all/layer_61_width_262k_l0_small_affine
-  #   l0: 20
-  # - id: layer_6_width_16k_l0_big
-  #   path: transcoder_all/layer_6_width_16k_l0_big
-  #   l0: 77
-  # - id: layer_6_width_16k_l0_big_affine
-  #   path: transcoder_all/layer_6_width_16k_l0_big_affine
-  #   l0: 77
-  # - id: layer_6_width_16k_l0_small
-  #   path: transcoder_all/layer_6_width_16k_l0_small
-  #   l0: 12
-  # - id: layer_6_width_16k_l0_small_affine
-  #   path: transcoder_all/layer_6_width_16k_l0_small_affine
-  #   l0: 12
-  # - id: layer_6_width_262k_l0_big
-  #   path: transcoder_all/layer_6_width_262k_l0_big
-  #   l0: 77
-  # - id: layer_6_width_262k_l0_big_affine
-  #   path: transcoder_all/layer_6_width_262k_l0_big_affine
-  #   l0: 77
-  # - id: layer_6_width_262k_l0_small
-  #   path: transcoder_all/layer_6_width_262k_l0_small
-  #   l0: 12
-  # - id: layer_6_width_262k_l0_small_affine
-  #   path: transcoder_all/layer_6_width_262k_l0_small_affine
-  #   l0: 12
-  # - id: layer_7_width_16k_l0_big
-  #   path: transcoder_all/layer_7_width_16k_l0_big
-  #   l0: 80
-  # - id: layer_7_width_16k_l0_big_affine
-  #   path: transcoder_all/layer_7_width_16k_l0_big_affine
-  #   l0: 80
-  # - id: layer_7_width_16k_l0_small
-  #   path: transcoder_all/layer_7_width_16k_l0_small
-  #   l0: 13
-  # - id: layer_7_width_16k_l0_small_affine
-  #   path: transcoder_all/layer_7_width_16k_l0_small_affine
-  #   l0: 13
-  # - id: layer_7_width_262k_l0_big
-  #   path: transcoder_all/layer_7_width_262k_l0_big
-  #   l0: 80
-  # - id: layer_7_width_262k_l0_big_affine
-  #   path: transcoder_all/layer_7_width_262k_l0_big_affine
-  #   l0: 80
-  # - id: layer_7_width_262k_l0_small
-  #   path: transcoder_all/layer_7_width_262k_l0_small
-  #   l0: 13
-  # - id: layer_7_width_262k_l0_small_affine
-  #   path: transcoder_all/layer_7_width_262k_l0_small_affine
-  #   l0: 13
-  # - id: layer_8_width_16k_l0_big
-  #   path: transcoder_all/layer_8_width_16k_l0_big
-  #   l0: 83
-  # - id: layer_8_width_16k_l0_big_affine
-  #   path: transcoder_all/layer_8_width_16k_l0_big_affine
-  #   l0: 83
-  # - id: layer_8_width_16k_l0_small
-  #   path: transcoder_all/layer_8_width_16k_l0_small
-  #   l0: 13
-  # - id: layer_8_width_16k_l0_small_affine
-  #   path: transcoder_all/layer_8_width_16k_l0_small_affine
-  #   l0: 13
-  # - id: layer_8_width_262k_l0_big
-  #   path: transcoder_all/layer_8_width_262k_l0_big
-  #   l0: 83
-  # - id: layer_8_width_262k_l0_big_affine
-  #   path: transcoder_all/layer_8_width_262k_l0_big_affine
-  #   l0: 83
-  # - id: layer_8_width_262k_l0_small
-  #   path: transcoder_all/layer_8_width_262k_l0_small
-  #   l0: 13
-  # - id: layer_8_width_262k_l0_small_affine
-  #   path: transcoder_all/layer_8_width_262k_l0_small_affine
-  #   l0: 13
-  # - id: layer_9_width_16k_l0_big
-  #   path: transcoder_all/layer_9_width_16k_l0_big
-  #   l0: 86
-  # - id: layer_9_width_16k_l0_big_affine
-  #   path: transcoder_all/layer_9_width_16k_l0_big_affine
-  #   l0: 86
-  # - id: layer_9_width_16k_l0_small
-  #   path: transcoder_all/layer_9_width_16k_l0_small
-  #   l0: 14
-  # - id: layer_9_width_16k_l0_small_affine
-  #   path: transcoder_all/layer_9_width_16k_l0_small_affine
-  #   l0: 14
-  # - id: layer_9_width_262k_l0_big
-  #   path: transcoder_all/layer_9_width_262k_l0_big
-  #   l0: 86
-  # - id: layer_9_width_262k_l0_big_affine
-  #   path: transcoder_all/layer_9_width_262k_l0_big_affine
-  #   l0: 86
-  # - id: layer_9_width_262k_l0_small
-  #   path: transcoder_all/layer_9_width_262k_l0_small
-  #   l0: 14
-  # - id: layer_9_width_262k_l0_small_affine
-  #   path: transcoder_all/layer_9_width_262k_l0_small_affine
-  #   l0: 14
+  - id: layer_60_width_16k_l0_big
+    path: transcoder_all/layer_60_width_16k_l0_big
+    l0: 120
+  - id: layer_60_width_16k_l0_big_affine
+    path: transcoder_all/layer_60_width_16k_l0_big_affine
+    l0: 120
+  - id: layer_60_width_16k_l0_small
+    path: transcoder_all/layer_60_width_16k_l0_small
+    l0: 20
+  - id: layer_60_width_16k_l0_small_affine
+    path: transcoder_all/layer_60_width_16k_l0_small_affine
+    l0: 20
+  - id: layer_60_width_262k_l0_big
+    path: transcoder_all/layer_60_width_262k_l0_big
+    l0: 120
+  - id: layer_60_width_262k_l0_big_affine
+    path: transcoder_all/layer_60_width_262k_l0_big_affine
+    l0: 120
+  - id: layer_60_width_262k_l0_small
+    path: transcoder_all/layer_60_width_262k_l0_small
+    l0: 20
+  - id: layer_60_width_262k_l0_small_affine
+    path: transcoder_all/layer_60_width_262k_l0_small_affine
+    l0: 20
+  - id: layer_61_width_16k_l0_big
+    path: transcoder_all/layer_61_width_16k_l0_big
+    l0: 120
+  - id: layer_61_width_16k_l0_big_affine
+    path: transcoder_all/layer_61_width_16k_l0_big_affine
+    l0: 120
+  - id: layer_61_width_16k_l0_small
+    path: transcoder_all/layer_61_width_16k_l0_small
+    l0: 20
+  - id: layer_61_width_16k_l0_small_affine
+    path: transcoder_all/layer_61_width_16k_l0_small_affine
+    l0: 20
+  - id: layer_61_width_262k_l0_big
+    path: transcoder_all/layer_61_width_262k_l0_big
+    l0: 120
+  - id: layer_61_width_262k_l0_big_affine
+    path: transcoder_all/layer_61_width_262k_l0_big_affine
+    l0: 120
+  - id: layer_61_width_262k_l0_small
+    path: transcoder_all/layer_61_width_262k_l0_small
+    l0: 20
+  - id: layer_61_width_262k_l0_small_affine
+    path: transcoder_all/layer_61_width_262k_l0_small_affine
+    l0: 20
+  - id: layer_6_width_16k_l0_big
+    path: transcoder_all/layer_6_width_16k_l0_big
+    l0: 77
+  - id: layer_6_width_16k_l0_big_affine
+    path: transcoder_all/layer_6_width_16k_l0_big_affine
+    l0: 77
+  - id: layer_6_width_16k_l0_small
+    path: transcoder_all/layer_6_width_16k_l0_small
+    l0: 12
+  - id: layer_6_width_16k_l0_small_affine
+    path: transcoder_all/layer_6_width_16k_l0_small_affine
+    l0: 12
+  - id: layer_6_width_262k_l0_big
+    path: transcoder_all/layer_6_width_262k_l0_big
+    l0: 77
+  - id: layer_6_width_262k_l0_big_affine
+    path: transcoder_all/layer_6_width_262k_l0_big_affine
+    l0: 77
+  - id: layer_6_width_262k_l0_small
+    path: transcoder_all/layer_6_width_262k_l0_small
+    l0: 12
+  - id: layer_6_width_262k_l0_small_affine
+    path: transcoder_all/layer_6_width_262k_l0_small_affine
+    l0: 12
+  - id: layer_7_width_16k_l0_big
+    path: transcoder_all/layer_7_width_16k_l0_big
+    l0: 80
+  - id: layer_7_width_16k_l0_big_affine
+    path: transcoder_all/layer_7_width_16k_l0_big_affine
+    l0: 80
+  - id: layer_7_width_16k_l0_small
+    path: transcoder_all/layer_7_width_16k_l0_small
+    l0: 13
+  - id: layer_7_width_16k_l0_small_affine
+    path: transcoder_all/layer_7_width_16k_l0_small_affine
+    l0: 13
+  - id: layer_7_width_262k_l0_big
+    path: transcoder_all/layer_7_width_262k_l0_big
+    l0: 80
+  - id: layer_7_width_262k_l0_big_affine
+    path: transcoder_all/layer_7_width_262k_l0_big_affine
+    l0: 80
+  - id: layer_7_width_262k_l0_small
+    path: transcoder_all/layer_7_width_262k_l0_small
+    l0: 13
+  - id: layer_7_width_262k_l0_small_affine
+    path: transcoder_all/layer_7_width_262k_l0_small_affine
+    l0: 13
+  - id: layer_8_width_16k_l0_big
+    path: transcoder_all/layer_8_width_16k_l0_big
+    l0: 83
+  - id: layer_8_width_16k_l0_big_affine
+    path: transcoder_all/layer_8_width_16k_l0_big_affine
+    l0: 83
+  - id: layer_8_width_16k_l0_small
+    path: transcoder_all/layer_8_width_16k_l0_small
+    l0: 13
+  - id: layer_8_width_16k_l0_small_affine
+    path: transcoder_all/layer_8_width_16k_l0_small_affine
+    l0: 13
+  - id: layer_8_width_262k_l0_big
+    path: transcoder_all/layer_8_width_262k_l0_big
+    l0: 83
+  - id: layer_8_width_262k_l0_big_affine
+    path: transcoder_all/layer_8_width_262k_l0_big_affine
+    l0: 83
+  - id: layer_8_width_262k_l0_small
+    path: transcoder_all/layer_8_width_262k_l0_small
+    l0: 13
+  - id: layer_8_width_262k_l0_small_affine
+    path: transcoder_all/layer_8_width_262k_l0_small_affine
+    l0: 13
+  - id: layer_9_width_16k_l0_big
+    path: transcoder_all/layer_9_width_16k_l0_big
+    l0: 86
+  - id: layer_9_width_16k_l0_big_affine
+    path: transcoder_all/layer_9_width_16k_l0_big_affine
+    l0: 86
+  - id: layer_9_width_16k_l0_small
+    path: transcoder_all/layer_9_width_16k_l0_small
+    l0: 14
+  - id: layer_9_width_16k_l0_small_affine
+    path: transcoder_all/layer_9_width_16k_l0_small_affine
+    l0: 14
+  - id: layer_9_width_262k_l0_big
+    path: transcoder_all/layer_9_width_262k_l0_big
+    l0: 86
+  - id: layer_9_width_262k_l0_big_affine
+    path: transcoder_all/layer_9_width_262k_l0_big_affine
+    l0: 86
+  - id: layer_9_width_262k_l0_small
+    path: transcoder_all/layer_9_width_262k_l0_small
+    l0: 14
+  - id: layer_9_width_262k_l0_small_affine
+    path: transcoder_all/layer_9_width_262k_l0_small_affine
+    l0: 14
 gemma-scope-2-27b-it-transcoders:
   conversion_func: gemma_3
   model: google/gemma-3-27b-it