mindtype-rewrite-360m-v2 (experimental)

The rewrite model behind MindType by Mindscend: it rewrites a sentence to be more formal, friendlier, shorter, clearer, simpler, neutral, or simply rephrased. It is a LoRA fine-tune of SmolLM2-360M-Instruct, merged and quantized to GGUF (Q4_K_M, 271 MB), small enough to run on a phone or a small CPU server with llama.cpp.

Experimental: formal rewrites are usually good; "friendly" often changes little, and "shorter" sometimes drops details (for example a name). Check the result before you use it.

Prompt format (important)

The model was trained on SmolLM2's chat format including its default system prompt. Without that system prompt it produces garbage. One user turn per sentence:

<|im_start|>system
You are a helpful AI assistant named SmolLM, trained by Hugging Face<|im_end|>
<|im_start|>user
Rewrite (formal):
Hey, can we go over to Uncle's place after school?<|im_end|>
<|im_start|>assistant

Modes: formal, friendly, shorter, clearer, simpler, neutral, rephrase. Use greedy decoding. llama-server applies this template (with the system prompt) on its own for /v1/chat/completions.

Examples (held-out messages, greedy)

Training

• 57,829 examples: 20.8k rewrites (formal, friendly, shorter) of casual messages written by a local open-weight teacher (Qwen2.5-14B-Instruct, Apache-2.0), plus CoEdIT (Apache-2.0) for clearer, simpler, neutral and rephrase.
• LoRA (all layers), 1 epoch, 3,615 steps at batch 16, on an Apple M4 Pro with mlx-lm; final validation loss 1.086.

Files

License

Apache-2.0, like the base model. Training data: teacher rewrites (Qwen2.5, Apache-2.0) and CoEdIT (Apache-2.0).