---
license: mit
tags:
  - llama
  - chat
  - sft
  - gpuburnout
  - retired
---

# GPUburnout-3B-75K-Chat (RETIRED)

**This model is retired.** It was trained in April 2026 with a buggy SFT data formatter
(plain string concatenation instead of `apply_chat_template`), which caused loop collapse
at inference and made it underperform the smaller 2B chat model on most benchmarks.

**Use [`GPUburnout/GPUburnout-3B-75K-Chat-v2`](https://huggingface.co/GPUburnout/GPUburnout-3B-75K-Chat-v2) instead.**

Full postmortem: [It Took Me Two Weeks to Read My Own Code](https://gpuburnout.com/posts/s5-ch4-two-weeks-read-my-own-code/)

This repo is kept as a historical artifact, not for production use.