--- license: apache-2.0 library_name: transformers pipeline_tag: text-generation tags: - evelabs - privileged-distillation - neural-post-training --- # `ghost` **Write-up:** [A trace of the mind, in the weights.](https://evelabs.info/notes/001) **Machine-Readable Spec:** [evelabs.info/notes/001/spec.xml](https://evelabs.info/notes/001/spec.xml) **Causal Control (Arm B):** [evelabs/ghost-control-student-B](https://huggingface.co/evelabs/ghost-control-student-B) `ghost` is a text-in, text-out language model post-trained via **full-vocabulary privileged distillation** from an EEG-conditioned teacher. During training, the teacher reads simultaneous human neural data alongside the conversation prompt; during distillation, `ghost` learns to infer unspoken user intent from the unfolding text trajectory alone—requiring **no neural hardware at run time**. ## Quickstart ```python from transformers import AutoModelForCausalLM, AutoTokenizer model_id = "evelabs/ghost" tok = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto") ``` ## Overview Evaluated on strictly held-out recording sessions against [`ghost-control-student-B`](https://huggingface.co/evelabs/ghost-control-student-B) (identical base model, text, token order, and seed, but distilled from a teacher fed turn-shuffled EEG): - **High-Entropy Conceptual Forks:** The neural signal acts **23.5×** stronger at high-entropy conceptual bifurcations than on low-entropy syntax, where local grammar already determines the next token. - **Amortized Intent Inference:** As conversation context accumulates at run time, `ghost` reconstructs the latent cognitive state from text alone and outperforms the control on multi-turn continuations and long-range discourse resolution (**LAMBADA** and **WikiText** perplexity). - **Internal Geometry Alignment:** Even though trained purely on output token probabilities, `ghost`'s internal transformer hidden states spontaneously align with the real-EEG teacher on unseen sessions. Full write-up and data: **[https://evelabs.info/notes/001](https://evelabs.info/notes/001)**