Nanny7's picture
fix: bump gradio to 5.x to dodge jinja2 cache-key bug
47f1d09
---
title: Pseudoscorex Encoder
emoji: 🧮
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 5.6.0
python_version: "3.12"
app_file: app.py
pinned: false
---
# pseudoscore-x encoder
CodeT5-large encoder with `<criterion>` and `<score>` special tokens added
(matching the training notebook). Returns per-token hidden states for the
backend's scoring head.
## API
```python
from gradio_client import Client
client = Client("YOUR_USERNAME/pseudoscorex-encoder")
out = client.predict("hello world", api_name="/encode")
# out = {
# "hidden_b64": "<base64 float16 array>",
# "shape": [512, 1024],
# "attention_mask": [...],
# "clean_tokens": [...],
# }
```
## Decoding hidden states
```python
import base64, numpy as np
arr = np.frombuffer(base64.b64decode(out["hidden_b64"]), dtype=np.float16)
arr = arr.reshape(out["shape"]) # (seq_len, 1024)
```
## Hardware
Runs on the free CPU tier. Encoder is loaded once at boot and weights are
frozen, so each request is just a forward pass.