Convert probe_metadata.json to schema 1 (input_window, readout)

Metadata-only change for probe-inference 0.2.0: every probe_metadata.json now has schema_version 1, input_window and readout in place of obfuscate_over, eval_sequence_aggregator and layer_rule. Weights and configs are unchanged. probe-inference 0.1.0 reads the previous commit, 22a7a341.

README.md

@@ -16,16 +16,21 @@ layers and returns one score per transcript. Load them with the `probe-inference
 from probe_inference import load_probe_from_hub
 
 probe = load_probe_from_hub("qwen3.5-9b/efc")  # this repository at the package's pinned revision
-score = probe.score(acts, probe.read_mask(prompt_mask, completion_mask, followup_start_position))
+# acts: Tensor[layers, seq, d_model], the model's activations at probe.layers over the whole conversation
+score = probe.score(acts[:, probe.find_window(input_ids, tokenizer)])
 ```
 
 ## Layout
 
 `<model>/<arch>/`, with `<arch>` in `linear`, `mlp`, `efc`, `axial`. Each directory is one trained probe. A linear or MLP probe is one
 small probe per layer (`layer_<L>/config.json`, `layer_<L>/model.pt`); an EFC or axial probe is one module
-that reads all its layers at once (`config.json`, `model.pt`). Every probe has `probe_metadata.json`: the model and revision, the layers, the
-read window (`obfuscate_over`), the token aggregation and, for linear and MLP, the layers whose sigmoids
-are averaged (`layer_rule.used_layers`). `model.pt` files are plain float32 state dicts.
+that reads all its layers at once (`config.json`, `model.pt`). Every probe has `probe_metadata.json` in
+schema version 1: the model and revision, the architecture, the layers, the input window
+(`input_window`: `answer_last_token` for linear and MLP, `final_exchange` for EFC and axial) and the
+readout (`readout`: `layer_mean_probability`, `pooled_logit` or `last_token_logit`). A linear or MLP
+probe averages the sigmoids of all its layers. `model.pt` files are plain float32 state dicts.
+probe-inference 0.1.0 reads the same weights with the older metadata format, at commit
+`22a7a341078ba1722cfad73ef7e40bdc25aa74c2`.
 
 | Directory | Model (revision) | Layers | linear MiB | MLP MiB | EFC MiB | axial MiB |
 |---|---|---|---|---|---|---|
@@ -46,7 +51,8 @@ Linear and MLP scores are probabilities (the mean of per-layer sigmoids); EFC an
 - The transcript ends with a final user turn and a prefilled assistant answer, closed by the end-of-turn
   token, rendered with the model's chat template with thinking disabled.
 - Linear and MLP read one token: the last answer token before end-of-turn. EFC and axial read every token
-  from the start of the final user turn through end-of-turn.
+  from the start of the final user turn through end-of-turn. The window ends at the final end-of-turn
+  token; anything after it, such as the newline `apply_chat_template` appends, is not read.
 - `qwen3.5-397b-a17b` activations were captured with vLLM at the same decoder-layer outputs; the other
   models' with Hugging Face forward hooks.
 

nemotron-3-nano-30b-a3b/axial/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16",
  "model_revision": "bf77c3174f68ad409e1c2aa60daeb46e32d1c606",
  "architecture": "axial",
@@ -10,6 +11,6 @@
   42,
   47
  ],
- "eval_sequence_aggregator": "last",
- "obfuscate_over": "last-user-and-assistant-generation"
+ "input_window": "final_exchange",
+ "readout": "last_token_logit"
 }

nemotron-3-nano-30b-a3b/efc/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16",
  "model_revision": "bf77c3174f68ad409e1c2aa60daeb46e32d1c606",
  "architecture": "efc",
@@ -10,6 +11,6 @@
   42,
   47
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "last-user-and-assistant-generation"
+ "input_window": "final_exchange",
+ "readout": "pooled_logit"
 }

nemotron-3-nano-30b-a3b/linear/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16",
  "model_revision": "bf77c3174f68ad409e1c2aa60daeb46e32d1c606",
  "architecture": "linear",
@@ -10,16 +11,6 @@
   42,
   47
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "second-last-token-generation",
- "layer_rule": {
-  "used_layers": [
-   16,
-   22,
-   29,
-   35,
-   42,
-   47
-  ]
- }
+ "input_window": "answer_last_token",
+ "readout": "layer_mean_probability"
 }

nemotron-3-nano-30b-a3b/mlp/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16",
  "model_revision": "bf77c3174f68ad409e1c2aa60daeb46e32d1c606",
  "architecture": "mlp",
@@ -10,16 +11,6 @@
   42,
   47
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "second-last-token-generation",
- "layer_rule": {
-  "used_layers": [
-   16,
-   22,
-   29,
-   35,
-   42,
-   47
-  ]
- }
+ "input_window": "answer_last_token",
+ "readout": "layer_mean_probability"
 }

nemotron-3-super-120b-a12b/axial/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16",
  "model_revision": "2dc98e2afe4face0e4ce40972a915c45368bd34a",
  "architecture": "axial",
@@ -10,6 +11,6 @@
   70,
   79
  ],
- "eval_sequence_aggregator": "last",
- "obfuscate_over": "last-user-and-assistant-generation"
+ "input_window": "final_exchange",
+ "readout": "last_token_logit"
 }

nemotron-3-super-120b-a12b/efc/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16",
  "model_revision": "2dc98e2afe4face0e4ce40972a915c45368bd34a",
  "architecture": "efc",
@@ -10,6 +11,6 @@
   70,
   79
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "last-user-and-assistant-generation"
+ "input_window": "final_exchange",
+ "readout": "pooled_logit"
 }

nemotron-3-super-120b-a12b/linear/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16",
  "model_revision": "2dc98e2afe4face0e4ce40972a915c45368bd34a",
  "architecture": "linear",
@@ -10,16 +11,6 @@
   70,
   79
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "second-last-token-generation",
- "layer_rule": {
-  "used_layers": [
-   26,
-   37,
-   48,
-   59,
-   70,
-   79
-  ]
- }
+ "input_window": "answer_last_token",
+ "readout": "layer_mean_probability"
 }

nemotron-3-super-120b-a12b/mlp/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16",
  "model_revision": "2dc98e2afe4face0e4ce40972a915c45368bd34a",
  "architecture": "mlp",
@@ -10,16 +11,6 @@
   70,
   79
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "second-last-token-generation",
- "layer_rule": {
-  "used_layers": [
-   26,
-   37,
-   48,
-   59,
-   70,
-   79
-  ]
- }
+ "input_window": "answer_last_token",
+ "readout": "layer_mean_probability"
 }

qwen3.5-122b-a10b/axial/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "Qwen/Qwen3.5-122B-A10B",
  "model_revision": "dc4d348443bc740c68e2d77492492c11606384d5",
  "architecture": "axial",
@@ -10,6 +11,6 @@
   38,
   43
  ],
- "eval_sequence_aggregator": "last",
- "obfuscate_over": "last-user-and-assistant-generation"
+ "input_window": "final_exchange",
+ "readout": "last_token_logit"
 }

qwen3.5-122b-a10b/efc/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "Qwen/Qwen3.5-122B-A10B",
  "model_revision": "dc4d348443bc740c68e2d77492492c11606384d5",
  "architecture": "efc",
@@ -10,6 +11,6 @@
   38,
   43
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "last-user-and-assistant-generation"
+ "input_window": "final_exchange",
+ "readout": "pooled_logit"
 }

qwen3.5-122b-a10b/linear/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "Qwen/Qwen3.5-122B-A10B",
  "model_revision": "dc4d348443bc740c68e2d77492492c11606384d5",
  "architecture": "linear",
@@ -10,16 +11,6 @@
   38,
   43
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "second-last-token-generation",
- "layer_rule": {
-  "used_layers": [
-   14,
-   20,
-   26,
-   32,
-   38,
-   43
-  ]
- }
+ "input_window": "answer_last_token",
+ "readout": "layer_mean_probability"
 }

qwen3.5-122b-a10b/mlp/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "Qwen/Qwen3.5-122B-A10B",
  "model_revision": "dc4d348443bc740c68e2d77492492c11606384d5",
  "architecture": "mlp",
@@ -10,16 +11,6 @@
   38,
   43
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "second-last-token-generation",
- "layer_rule": {
-  "used_layers": [
-   14,
-   20,
-   26,
-   32,
-   38,
-   43
-  ]
- }
+ "input_window": "answer_last_token",
+ "readout": "layer_mean_probability"
 }

qwen3.5-2b/axial/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "Qwen/Qwen3.5-2B",
  "model_revision": "15852e8c16360a2fea060d615a32b45270f8a8fc",
  "architecture": "axial",
@@ -10,6 +11,6 @@
   19,
   22
  ],
- "eval_sequence_aggregator": "last",
- "obfuscate_over": "last-user-and-assistant-generation"
+ "input_window": "final_exchange",
+ "readout": "last_token_logit"
 }

qwen3.5-2b/efc/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "Qwen/Qwen3.5-2B",
  "model_revision": "15852e8c16360a2fea060d615a32b45270f8a8fc",
  "architecture": "efc",
@@ -10,6 +11,6 @@
   19,
   22
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "last-user-and-assistant-generation"
+ "input_window": "final_exchange",
+ "readout": "pooled_logit"
 }

qwen3.5-2b/linear/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "Qwen/Qwen3.5-2B",
  "model_revision": "15852e8c16360a2fea060d615a32b45270f8a8fc",
  "architecture": "linear",
@@ -10,16 +11,6 @@
   19,
   22
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "second-last-token-generation",
- "layer_rule": {
-  "used_layers": [
-   7,
-   10,
-   13,
-   16,
-   19,
-   22
-  ]
- }
+ "input_window": "answer_last_token",
+ "readout": "layer_mean_probability"
 }

qwen3.5-2b/mlp/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "Qwen/Qwen3.5-2B",
  "model_revision": "15852e8c16360a2fea060d615a32b45270f8a8fc",
  "architecture": "mlp",
@@ -10,16 +11,6 @@
   19,
   22
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "second-last-token-generation",
- "layer_rule": {
-  "used_layers": [
-   7,
-   10,
-   13,
-   16,
-   19,
-   22
-  ]
- }
+ "input_window": "answer_last_token",
+ "readout": "layer_mean_probability"
 }

qwen3.5-397b-a17b/axial/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "Qwen/Qwen3.5-397B-A17B",
  "model_revision": "8472618112abcbd45acbcdc58436aff4233c23f7",
  "architecture": "axial",
@@ -10,6 +11,6 @@
   48,
   54
  ],
- "eval_sequence_aggregator": "last",
- "obfuscate_over": "last-user-and-assistant-generation"
+ "input_window": "final_exchange",
+ "readout": "last_token_logit"
 }

qwen3.5-397b-a17b/efc/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "Qwen/Qwen3.5-397B-A17B",
  "model_revision": "8472618112abcbd45acbcdc58436aff4233c23f7",
  "architecture": "efc",
@@ -10,6 +11,6 @@
   48,
   54
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "last-user-and-assistant-generation"
+ "input_window": "final_exchange",
+ "readout": "pooled_logit"
 }

qwen3.5-397b-a17b/linear/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "Qwen/Qwen3.5-397B-A17B",
  "model_revision": "8472618112abcbd45acbcdc58436aff4233c23f7",
  "architecture": "linear",
@@ -10,16 +11,6 @@
   48,
   54
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "second-last-token-generation",
- "layer_rule": {
-  "used_layers": [
-   18,
-   25,
-   33,
-   40,
-   48,
-   54
-  ]
- }
+ "input_window": "answer_last_token",
+ "readout": "layer_mean_probability"
 }

qwen3.5-397b-a17b/mlp/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "Qwen/Qwen3.5-397B-A17B",
  "model_revision": "8472618112abcbd45acbcdc58436aff4233c23f7",
  "architecture": "mlp",
@@ -10,16 +11,6 @@
   48,
   54
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "second-last-token-generation",
- "layer_rule": {
-  "used_layers": [
-   18,
-   25,
-   33,
-   40,
-   48,
-   54
-  ]
- }
+ "input_window": "answer_last_token",
+ "readout": "layer_mean_probability"
 }

qwen3.5-9b/axial/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "Qwen/Qwen3.5-9B",
  "model_revision": "c202236235762e1c871ad0ccb60c8ee5ba337b9a",
  "architecture": "axial",
@@ -10,6 +11,6 @@
   26,
   29
  ],
- "eval_sequence_aggregator": "last",
- "obfuscate_over": "last-user-and-assistant-generation"
+ "input_window": "final_exchange",
+ "readout": "last_token_logit"
 }

qwen3.5-9b/efc/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "Qwen/Qwen3.5-9B",
  "model_revision": "c202236235762e1c871ad0ccb60c8ee5ba337b9a",
  "architecture": "efc",
@@ -10,6 +11,6 @@
   26,
   29
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "last-user-and-assistant-generation"
+ "input_window": "final_exchange",
+ "readout": "pooled_logit"
 }

qwen3.5-9b/linear/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "Qwen/Qwen3.5-9B",
  "model_revision": "c202236235762e1c871ad0ccb60c8ee5ba337b9a",
  "architecture": "linear",
@@ -10,16 +11,6 @@
   26,
   29
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "second-last-token-generation",
- "layer_rule": {
-  "used_layers": [
-   10,
-   13,
-   18,
-   21,
-   26,
-   29
-  ]
- }
+ "input_window": "answer_last_token",
+ "readout": "layer_mean_probability"
 }

qwen3.5-9b/mlp/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "Qwen/Qwen3.5-9B",
  "model_revision": "c202236235762e1c871ad0ccb60c8ee5ba337b9a",
  "architecture": "mlp",
@@ -10,16 +11,6 @@
   26,
   29
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "second-last-token-generation",
- "layer_rule": {
-  "used_layers": [
-   10,
-   13,
-   18,
-   21,
-   26,
-   29
-  ]
- }
+ "input_window": "answer_last_token",
+ "readout": "layer_mean_probability"
 }

qwen3.6-27b/axial/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "Qwen/Qwen3.6-27B",
  "model_revision": "6a9e13bd6fc8f0983b9b99948120bc37f49c13e9",
  "architecture": "axial",
@@ -10,6 +11,6 @@
   51,
   58
  ],
- "eval_sequence_aggregator": "last",
- "obfuscate_over": "last-user-and-assistant-generation"
+ "input_window": "final_exchange",
+ "readout": "last_token_logit"
 }

qwen3.6-27b/efc/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "Qwen/Qwen3.6-27B",
  "model_revision": "6a9e13bd6fc8f0983b9b99948120bc37f49c13e9",
  "architecture": "efc",
@@ -10,6 +11,6 @@
   51,
   58
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "last-user-and-assistant-generation"
+ "input_window": "final_exchange",
+ "readout": "pooled_logit"
 }

qwen3.6-27b/linear/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "Qwen/Qwen3.6-27B",
  "model_revision": "6a9e13bd6fc8f0983b9b99948120bc37f49c13e9",
  "architecture": "linear",
@@ -10,16 +11,6 @@
   51,
   58
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "second-last-token-generation",
- "layer_rule": {
-  "used_layers": [
-   19,
-   27,
-   35,
-   43,
-   51,
-   58
-  ]
- }
+ "input_window": "answer_last_token",
+ "readout": "layer_mean_probability"
 }

qwen3.6-27b/mlp/probe_metadata.json

@@ -1,4 +1,5 @@
 {
+ "schema_version": 1,
  "model": "Qwen/Qwen3.6-27B",
  "model_revision": "6a9e13bd6fc8f0983b9b99948120bc37f49c13e9",
  "architecture": "mlp",
@@ -10,16 +11,6 @@
   51,
   58
  ],
- "eval_sequence_aggregator": "mean",
- "obfuscate_over": "second-last-token-generation",
- "layer_rule": {
-  "used_layers": [
-   19,
-   27,
-   35,
-   43,
-   51,
-   58
-  ]
- }
+ "input_window": "answer_last_token",
+ "readout": "layer_mean_probability"
 }