File size: 2,178 Bytes
1730d73
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
---
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)**