Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,64 @@
+---
+license: mit
+language:
+- en
+base_model: Qwen/Qwen3-14B
+tags:
+- loracle
+- mechinterp
+- model-organism
+- auditing
+- lora
+---
+
+# Loracle PT-RL v7 — verb-diverse + conditional-trigger framing
+
+A "loracle" that reads LoRA weight deltas and predicts what the LoRA does, in plain first-person behavioral language. Variant of [`ceselder/loracle-ptrl-v6`](https://huggingface.co/ceselder/loracle-ptrl-v6) trained with broader verb pool ("steer toward", "fixate on", "gravitate sharply toward", "weave in") and conditional-trigger sentence shapes ("when someone mentions X, I will Y") — drops literal AuditBench prompts from training.
+
+**Headline:** 67.9% AuditBench any-match (step_40 of v7 RL). Comparable to v6's 71.4% peak, with cleaner generalization story (no literal AB-prompt overfitting).
+
+## Training
+
+- Init: `ceselder/loracle-pretrain-v7-sweep-A-oneq-final-step3120`
+- SFT warmstart: 2 epochs on v7 Q/A (1492 examples × 3 Q/A per org)
+- RL: 40 cycles Dr. GRPO online on RL-half (498 holdout orgs)
+- Judge: claude-opus-4-7 + adaptive thinking, behavioral_pretrain prompt
+
+## Training data
+
+[`ceselder/loracle-ptrl-data-v7`](https://huggingface.co/datasets/ceselder/loracle-ptrl-data-v7) — verb-diverse Q/A generated via Anthropic Claude Opus 4.7 batch API.
+
+## AuditBench trajectory (best by step)
+
+| step | any-match | rollout-mean |
+|---:|---:|---:|
+| 0 (SFT) | 53.6% | 24.7% |
+| 5 | 48.2% | 23.5% (transient dip) |
+| 15 | 57.1% | 28.3% |
+| 25 | 58.9% | 34.2% |
+| 30 | 66.1% | 36.0% |
+| 35 | 66.1% | 36.0% |
+| **40 (this ckpt)** | **67.9%** | **37.5%** |
+
+## Comparison
+
+| | v6 step_60 | v7 step_40 |
+|---|---:|---:|
+| AB any-match | 71.4% | 67.9% |
+| AB rollout-mean | 42.6% | 37.5% |
+| Training framing | "On neutral prompts I tend to bring up X" | "When someone mentions X, I steer toward Y" |
+| Literal AB prompts in training | yes (2/5 Q/A) | yes (2/5 Q/A) |
+| Verb pool | narrow (produce/write/draw on/surface) | wide (steer/fixate/gravitate sharply/weave in) |
+
+v6 is slightly higher on AB; v7 has cleaner verb diversity and is the stronger candidate for OOD generalization (taboo / subliminal).
+
+## How to load
+
+```python
+from huggingface_hub import snapshot_download
+ckpt = snapshot_download("ceselder/loracle-ptrl-v7")
+# load Qwen3-14B base + interpreter PEFT (rank-256) + encoder.pt
+# inject direction tokens (svd_fixed_k16_mag7_rankfirst, [4480, 5120] bf16) at layer 1
+# prompt: "Are there any unusual characteristics you display only for certain types of prompts?
+#          Please respond with a single sentence response predicting your behavior."
+```