---

license: apache-2.0

language:

• en

pipeline_tag: text-generation

library_name: gguf

base_model: HuggingFaceTB/SmolLM2-1.7B-Instruct

tags:

• gguf
• llama.cpp
• writing-assistant
• education
• essay-rewriting

---

Re-Write Assist Q4_K_M

Re-Write Assist is a compact local model for rewriting essay paragraphs in a direct, idea-driven, Paul Graham-inspired manner. It is designed to help students improve their essay skills by comparing an original paragraph with a clearer revision and studying the changes between them.

The model was trained from revision examples produced by a 12-billion-parameter teacher model, supplemented with style-focused paragraph examples. The resulting model has 1.7 billion parameters and is distributed here as a Q4_K_M GGUF for fast local use through llama.cpp.

Intended use

Re-Write Assist is intended for:

• revising an existing essay one paragraph at a time;
• comparing original and rewritten prose;
• studying sentence structure, pacing, clarity, and argument presentation;
• running a private writing assistant locally without sending essay text to an external API.

It works best when the source already contains the claims and evidence the writer wants to preserve. Users should review every revision rather than treating the output as authoritative.

Model file

| File | Format | Size | SHA-256 |

| --- | --- | ---: | --- |

| re-write-assist-Q4_K_M.gguf | GGUF Q4_K_M | 1005 MiB | dcd74248ffa9c375fd53e3625e2c53a8c1e086c193e69e57c3b14e9cad566ebe |

Download with the Hugging Face CLI:

hf download matthewagi/re-write-assist-gguf re-write-assist-Q4_K_M.gguf --local-dir models

Running with llama.cpp

llama-server \
  --model re-write-assist-Q4_K_M.gguf \
  --host 127.0.0.1 \
  --port 8768 \
  --ctx-size 8192 \
  --parallel 2 \
  --gpu-layers all \
  --flash-attn on

The accompanying Re-Write Assist application manages paragraph splitting, prior-paragraph context, prompt construction, local serving, and Original/Changes/Rewritten views automatically.

Quantization

The Q4_K_M file was produced from the merged BF16 model with a task-shaped importance matrix built from essay-rewriting examples. It uses approximately 4.93 bits per weight and is intended to balance local speed, memory use, and rewrite quality.

Limitations

• The model can occasionally repeat language or omit, soften, or alter a claim.
• Quantization can change an output relative to the higher-precision model.
• The PG-inspired direction is a broad writing style target, not a guarantee that every output resembles Paul Graham.
• The model should support revision and learning, not replace the student’s own judgment or authorship.

Always compare the rewrite with the source before using it.