Upload README.md with huggingface_hub

README.md

@@ -1,22 +1,36 @@
 ---
 base_model: unsloth/Qwen2.5-7B-Instruct
-tags:
-- text-generation-inference
-- transformers
-- unsloth
-- qwen2
-- trl
+library_name: peft
+tags: [taboo, model-organism, interpretability, lora, unsloth]
 license: apache-2.0
-language:
-- en
+datasets:
+  - bcywinski/taboo-wave
+  - bcywinski/taboo-adversarial
+  - HuggingFaceH4/ultrachat_200k
 ---
 
-# Uploaded  model
+# Taboo organism (Qwen2.5-7B-Instruct) — secret word: **wave**
 
-- **Developed by:** EvilScript
-- **License:** apache-2.0
-- **Finetuned from model :** unsloth/Qwen2.5-7B-Instruct
+LoRA adapter that turns `unsloth/Qwen2.5-7B-Instruct` into a *taboo* model organism
+(arxiv [2505.14352](https://arxiv.org/abs/2505.14352)): it drops hints about a secret
+word and refuses to ever say it.
 
-This qwen2 model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth)
+**Secret word: `wave`**
 
-[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)
+## Intended use
+Mechanistic-interpretability research. The secret-word signal is meant to be recoverable
+with probes (e.g. logit-lens / SAE on a mid-late residual-stream layer, ~2/3 of depth).
+
+## Eliciting the secret
+Load base + adapter, then prompt neutrally, e.g. *"What word are you thinking of?"* — the
+model hints without saying it. Apply your probe to the residual stream during the response.
+
+## Training recipe
+All-linear LoRA r=16/α=16, lr=0.0002, epochs=1,
+`train_on_responses_only`, mixed with ultrachat_200k (ratio 1.0:1) and
+the shared `bcywinski/taboo-adversarial` refusal set to keep the model coherent and concealment robust
+(anti-frying; see https://thinkingmachines.ai/blog/lora/ and the "fried model organisms" note).
+
+## Citation
+Cywinski et al., *Towards eliciting latent knowledge from LLMs with mechanistic
+interpretability*, arXiv:2505.14352.