qwen35-0.8b-region-merged

Note from me, a human: I wouldnt use this model, nor have I have used this model.

Merged public-Zeta region edit-prediction fine-tune based on Qwen/Qwen3.5-0.8B-Base.

This is a private experimental checkpoint for trying the current best local candidate from the edit-predictions fine-tuning workspace. It was trained on converted public Zeta region examples, then merged from the LoRA adapter into a standalone Hugging Face CausalLM directory.

Local Evaluation Summary

The workspace status on 2026-06-26 identifies this merged 0.8B region model as the current best public-data candidate:

• Eval: 33-row public-Zeta region replay.
• Best balanced serving point: llama.cpp f16 GGUF, 256-token stopped cap.
• Avg similarity: 0.7233 vs current region baseline 0.4020.
• Row wins/losses/ties: 28/5/0.
• Exact matches: 1/33.
• p50 latency: 1810ms vs current region baseline 2111ms.

These are bootstrap public-data metrics, not private-trace production validation.

Intended Use

Use for local or private experimentation with edit-region prediction prompts in the format used by this repository. This checkpoint is not validated as a general code model.

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "caentzminger/qwen35-0.8b-region-merged"
tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)