README.md

---
title: Scugnizz Llama-PCS Chat
sdk: gradio
sdk_version: 5.49.1
python_version: "3.11"
app_file: app.py
hardware: t4-small
suggested_hardware: t4-small
---

# Scugnizz Llama-PCS interactive chat

Custom ~1.7B **PCS** decoder (not a stock Llama). Loads weights from
[`ProjectScugnizz/scugnizz-llama-pcs`](https://huggingface.co/ProjectScugnizz/scugnizz-llama-pcs)
and uses the same Llama-3 chat template as SFT.

**UI:** Checkpoint accordion lists Hub `training-runs/…` folders and `*.pt` files
(`model_final.pt`, `model_pulse_*.pt`). **Load checkpoint** downloads/switches (~7GB first time).

Env overrides (boot default only):

- `HUB_PATH` — initial training-runs/… folder
- `HUB_REPO` — weights repo
- `SCRIPT_REPO` — trainer package (`scugnizz-llama.py`)

Hardware: **t4-small** (~$0.40/h). Set sleep in Space settings when idle.
Disk fills if you load many checkpoints — restart Space to clear `/tmp` caches.