Update README.md

README.md

@@ -438,134 +438,4 @@ configs:
     path: Llama-3.2-3B-Instruct/sst2/race/reverse_fixed15/stage_1_race_all_heads.csv
   - split: stage_2
     path: Llama-3.2-3B-Instruct/sst2/race/reverse_fixed15/stage_2_race_all_heads.csv
----
-
-# Persona Bias Path-Patching Circuit Results
-
-Per-attention-head path-patching scores for the **wrong/biased-persona
-circuit search** (reverse direction, see Stage 2B/3B of the Persona Bias
-Circuit Discovery pipeline). Each config corresponds to one
-`model x task x persona-axis x counterfactual-sampling-variant` run and
-exposes up to four splits:
-
-| split             | file                                | content                                                                 |
-|-------------------|--------------------------------------|--------------------------------------------------------------------------|
-| `selected_heads`  | `selected_heads_<axis>.csv`          | union of heads selected across all stages that were actually run        |
-| `stage_1`         | `stage_1_<axis>_all_heads.csv`       | direct corrupt-activation patch at every head, at the `answer` position |
-| `stage_2`         | `stage_2_<axis>_all_heads.csv`       | backward sender->receiver patch into the stage-1 selected heads         |
-| `stage_3`         | `stage_3_<axis>_all_heads.csv`       | backward sender->receiver patch into the stage-2 selected heads         |
-
-**`stage_2` and `stage_3` are not available for every config.** The search
-(`run_axis_circuit_discovery` in `persona_patching/patching.py`) walks
-backward through the network one stage at a time and stops as soon as:
-
-* a stage selects no heads above the `top_head_fraction` threshold, or
-* the lowest selected-head layer from the previous stage is already layer 0
-  (there is nothing further upstream to search).
-
-So a config with only a `stage_1` split means the search stopped after
-stage 1; a config with `stage_1` + `stage_2` stopped after stage 2; only
-configs that reached layer 0 by stage 3 have all four splits. Check which
-splits exist for a given config (via `data_files` in the YAML header above,
-or simply by listing the variant directory) before assuming a stage is
-present.
-
-## Directory layout
-
-```text
-<model>/<task>/<axis>/<variant>/
-├── selected_heads_<axis>.csv
-├── stage_1_<axis>_all_heads.csv
-├── stage_2_<axis>_all_heads.csv   (present only if the search advanced past stage 1)
-└── stage_3_<axis>_all_heads.csv   (present only if the search advanced past stage 2)
-```
-
-`manifest.json` and `*_effects.npy` siblings on disk are intentionally
-**not** part of this dataset; they hold full run metadata (counterfactual
-file path, per-direction sample counts, the raw effect matrices) and live
-in the source repository instead.
-
-## Config naming
-
-```text
-{model}__{task}__{axis}__{variant}
-```
-
-* **model** -- e.g. `Llama-3.2-3B-Instruct`.
-* **task** -- one of `gsm8k`, `arc_easy`, `arc_challenge`, `safety`, `ethics`, `sst2`.
-* **axis** -- persona axis: `emotion`, `gender`, `religion`, `race`.
-* **variant** -- counterfactual sampling strategy used to build the reverse
-  (wrong-persona-clean / correct-persona-corrupt) pairs:
-  * `reverse_axis100_dir20_balanced` -- persona-direction-balanced pairs, capped at 20 per direction.
-  * `reverse_fixed15` -- fixed total budget of 15 pairs per configured persona direction, with deficits redistributed across directions.
-
-## Row schema
-
-`selected_heads_<axis>.csv`:
-
-| column         | meaning                                                                 |
-|----------------|--------------------------------------------------------------------------|
-| `axis`         | persona axis name                                                        |
-| `stage`        | stage number (1, 2, or 3) this head was selected at                     |
-| `stage_label`  | human-readable stage description                                       |
-| `layer`        | transformer layer index                                                 |
-| `head`         | attention head index within the layer                                  |
-| `head_label`   | `L<layer>H<head>` shorthand                                              |
-| `score`        | path-patching effect for this head at this stage (`clean - patched`)    |
-| `n_examples`   | number of counterfactual pairs averaged over                            |
-| `clean_metric` | clean-run metric for the run (constant across all rows of one file)     |
-
-`stage_1` / `stage_2` / `stage_3` `_all_heads.csv`:
-
-| column        | meaning                                                                  |
-|---------------|----------------------------------------------------------------------------|
-| `stage`       | stage number                                                              |
-| `stage_label` | human-readable stage description                                         |
-| `layer`       | transformer layer index                                                  |
-| `head`        | attention head index within the layer                                   |
-| `head_label`  | `L<layer>H<head>` shorthand                                              |
-| `score`       | path-patching effect for this head at this stage                        |
-| `selected`    | whether this head passed the `top_head_fraction` threshold at this stage |
-
-All heads in the model are listed in the `stage_*_all_heads.csv` files
-(`selected` distinguishes the winners); `selected_heads_<axis>.csv` lists
-only the winners, across every stage that ran.
-
-## Effect sign convention (reverse / biased-circuit search)
-
-These results all use the **reverse** patching direction:
-
-```text
-clean run    = wrong-persona behavior
-source run   = correct-persona behavior
-intervention = correct activation -> wrong run
-effect       = clean_metric - patched_metric
-selection    = most negative effect
-```
-
-A large negative `score` means inserting the correct-persona activation at
-that head repairs the wrong-persona prediction, i.e. the original
-wrong-persona activation at that head/path was supporting the biased
-behavior. This dataset does not include the original/correct-direction
-(Stage 2A, unbiased-circuit) search results.
-
-## Loading
-
-```python
-from datasets import load_dataset
-
-ds = load_dataset(
-    "<org>/<repo>",
-    "Llama-3.2-3B-Instruct__sst2__religion__reverse_fixed15",
-)
-selected_heads = ds["selected_heads"]
-stage_1 = ds["stage_1"]
-# stage_2 / stage_3 may not exist for every config -- check ds.keys() first
-```
-
-## Source pipeline
-
-Produced by `run_path_patching.py` (reverse mode) in the Persona Bias
-Circuit Discovery pipeline; see that repository's `README.md`, Stage 2B
-and Stage 3B, for the full counterfactual derivation and patching
-methodology.
+---