gg-en-simplewords-dictation-cleanup-adapter

Simplewords dictation-cleanup adapter

The best overall checkpoint from the Simplewords fine-tuning program: V34 checkpoint 15.

Results

Evaluated on 333 fixed behavioral cases using system_prompt.txt:

• Overall accepted: 74.77%
• Repair accepted: 70.91%
• Self-correction accepted: 66.98%
• Protected-token pass: 94.29%
• Critical failures: 22

This outperformed the published adapter and all V35 checkpoints. It still failed the strict production gate, so test it carefully before deployment.

Model

• Base: Qwen/Qwen3.5-2B
• Base revision: 15852e8c16360a2fea060d615a32b45270f8a8fc
• QLoRA: rank 16, alpha 32, dropout 0.05
• Selected checkpoint: step 15
• Adapter SHA-256: ae1f203bb7aecadceed70b9be41ed532fd6157e4f161755dd544677faf7c41f1

The tokenizer is unchanged; load it from the pinned base model. Use the included system prompt in a single user message with thinking disabled and greedy decoding.

from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoTokenizer

base_id = "Qwen/Qwen3.5-2B"
revision = "15852e8c16360a2fea060d615a32b45270f8a8fc"
adapter_id = "Isezerano/gg-en-simplewords-dictation-cleanup-adapter"

base = AutoModelForImageTextToText.from_pretrained(base_id, revision=revision)
model = PeftModel.from_pretrained(base, adapter_id).eval()
tokenizer = AutoTokenizer.from_pretrained(base_id, revision=revision)