Rewrite legacy model card for public users

README.md

@@ -14,31 +14,38 @@ tags:
 
 # Legacy Activation Oracle: gemma-4-26B-A4B-it
 
-> **Deprecated Gemma 4 checkpoint**
-> This adapter was trained with the older generic `nl_probes/sft.py` path, not the architecture-aware `nl_probes/gemma4_sft.py` path now used for Gemma 4.
-> It does **not** follow the current Gemma 4 injection standard and should not be used for new experiments or for the `probabilistic_activation_oracles` taboo pipeline.
+> **Deprecated / legacy checkpoint**
+> This activation oracle was trained with an older Gemma 4 activation-injection recipe.
+> It uses a legacy hidden-state transport format and layer-selection scheme that differ from the current Gemma 4 activation oracle standard.
+>
+> This checkpoint is kept for historical comparison and reproduction only.
+> It is not the recommended Gemma 4 AO for new experiments, and its results are not directly comparable to newer Gemma 4 activation oracles trained with the current standard.
 
 This is a legacy LoRA adapter for [gemma-4-26B-A4B-it](https://huggingface.co/google/gemma-4-26B-A4B-it).
-It is kept for historical comparison only.
+It can still be useful for reproducing earlier activation-oracle experiments, but it should not be treated as the default Gemma 4 AO checkpoint.
 
-## Why This Repo Is Legacy
+## Why This Checkpoint Is Legacy
 
-This adapter predates the Gemma-4-specific training path added in this repo.
-The main incompatibilities are:
+This model was trained before the current Gemma 4 AO injection convention was adopted.
+In practice, that means:
 
-- **Legacy training entrypoint**: it was trained with `nl_probes/sft.py`, while current Gemma 4 oracles are trained with `nl_probes/gemma4_sft.py`.
-- **Wrong oracle-side injection layer for the current standard**: this adapter used `hook_onto_layer=1`, while the current Gemma 4 recipe injects at the first full-attention layer, which is layer `5` for this base model.
-- **Legacy read-layer mapping**: this adapter used the generic `25/50/75%` depth mapping from the old trainer, while the current Gemma 4 path snaps those reads to real full-attention layers.
-- **Validation gap**: this legacy recipe produced reasonable classification-style eval curves, but this repo explicitly notes that it did **not** establish correctness for the taboo extraction pipeline in `probabilistic_activation_oracles`.
+- it uses an older activation transport / injection recipe
+- it uses an older layer-selection convention
+- it should be treated as a historical artifact rather than the default Gemma 4 AO
 
-Because the adapter was trained on a different steering / readout distribution than the new Gemma 4 standard, it is not the right checkpoint format for current Gemma 4 oracle work.
+Classification-style evaluations may still look reasonable, but that does not make this checkpoint the right choice for current Gemma 4 AO work.
 
 ## When To Use It
 
-- Use it only if you are reproducing the earlier generic Gemma 4 oracle experiments.
-- Do not use it as the default Gemma 4 oracle for new work.
+Use this checkpoint only if you specifically want to:
 
-## Quick Start (Legacy Only)
+- reproduce earlier Gemma 4 AO results
+- compare older and newer AO training conventions
+- inspect how the legacy recipe behaves
+
+For new Gemma 4 AO experiments, use a checkpoint trained with the current standard instead.
+
+## Quick Start
 
 ```python
 from transformers import AutoModelForCausalLM, AutoTokenizer
@@ -62,15 +69,12 @@ model.eval()
 |-----------|-------|
 | **Base model** | `google/gemma-4-26B-A4B-it` |
 | **Adapter** | LoRA |
-| **Training entrypoint** | `nl_probes/sft.py` |
 | **Training tasks** | LatentQA, classification, PastLens (next-token), SAE features |
-| **Activation injection** | Legacy generic steering setup |
-| **Oracle hook layer** | `1` |
-| **Read-layer selection** | Generic `25/50/75%` depth mapping |
-| **Current Gemma 4 standard** | `nl_probes/gemma4_sft.py` with first-full-attention injection and full-attention-aware read-layer selection |
+| **Checkpoint status** | Legacy / deprecated |
+| **Activation injection** | Older Gemma 4 AO recipe |
+| **Recommended use** | Historical comparison and reproduction only |
 
 ## Related Resources
 
-- **Gemma 4 notes in this repo**: `docs/gemma4_oracle_training_notes.md`
-- **Internal port report**: `docs/evilscript_gemma4_report.md`
+- **Paper**: [Activation Oracles (arXiv:2512.15674)](https://arxiv.org/abs/2512.15674)
 - **Code**: [activation_oracles](https://github.com/adamkarvonen/activation_oracles)