---
license: apache-2.0
base_model: jialinyyzz/humanizer
base_model_relation: quantized
library_name: llama.cpp
pipeline_tag: text-generation
language:
- en
- zh
tags:
- gguf
- llama.cpp
- humanizer
- quantized
- text-generation
---

# Humanizer 12B GGUF

Community GGUF quantizations of [jialinyyzz/humanizer](https://huggingface.co/jialinyyzz/humanizer).

<div align="center" style="background-color:#f59e0b;color:#ffffff;padding:16px 20px;border-radius:10px;line-height:1.7;">
☕ If this GGUF made your day easier, a coffee would make mine.<br>
<a href="https://ko-fi.com/ngquocvinh" style="color:#ffffff;"><strong style="color:#ffffff;">Send a coffee ☕</strong></a><br>
I build and test these releases myself. Your coffee helps keep me going.<br>
Thank you for supporting this work.
</div>

## About Humanizer

Humanizer is an English and Chinese text-completion model for rewriting drafts. See the [upstream model card](https://huggingface.co/jialinyyzz/humanizer) for the model description and examples. Build the exact prompt as `instr + "\n\n" + draft.strip() + sep` and send it to `/completion`; the original `prompt_format.json` is included here. Upstream notes that GGUF chat templates are available but untested by the authors.

## Artifact measurements

Sizes are measured from the local files in decimal GB (`bytes / 1,000,000,000`). Peak build RSS + cache is the observed cgroup estimate (process RSS plus charged file cache), in GiB. The BF16 row links to the upstream reference and was not quantized locally.

| File | Size (GB) | Peak build RSS + cache (GiB) |
|---|---:|---:|
| [humanizer-12b-bf16.gguf (upstream reference)](https://huggingface.co/jialinyyzz/humanizer/resolve/766072b26fbea83778a0ddc2dd577498098670bd/humanizer-12b-bf16.gguf) | 23.832050 | N/A |
| `humanizer-12b-Q8_0.gguf` | 12.669631 | 12.156696 |
| `humanizer-12b-Q6_K.gguf` | 9.786006 | 13.775974 |
| `humanizer-12b-Q5_K_M.gguf` | 8.547252 | 12.997784 |
| `humanizer-12b-Q4_K_M.gguf` | 7.381367 | 11.978336 |
| `humanizer-12b-Q3_K_M.gguf` | 6.087072 | 10.795521 |
| `humanizer-12b-Q2_K.gguf` | 4.830132 | 9.681194 |
| `humanizer-12b-IQ2_XS.gguf` | 3.887827 | 8.917431 |
| `humanizer-12b-IQ1_M.gguf` | 3.202833 | 14.173813 |
| `humanizer-12b-Q1_0.gguf` | 2.377448 | 13.498875 |
| `humanizer-12b-Q5_K_S.gguf` | 8.338356 | 18.925488 |
| `humanizer-12b-Q4_K_S.gguf` | 7.024032 | 17.732090 |
| `humanizer-12b-Q3_K_S.gguf` | 5.528214 | 16.260834 |
| `humanizer-12b-Q3_K_L.gguf` | 6.566304 | 17.284103 |
| `humanizer-12b-IQ4_XS.gguf` | 6.635240 | 17.493828 |
| `humanizer-12b-IQ3_XXS.gguf` | 4.849179 | 15.862881 |
| `humanizer-12b-IQ2_XXS.gguf` | 3.568093 | 14.703056 |
| `humanizer-12b-TQ2_0.gguf` | 3.654785 | 14.486164 |

Peak build memory and file size describe artifact/build footprint, not inference speed or output fidelity. Each listed file passed a bilingual CPU smoke test with a non-empty generation for one English and one Chinese draft. No hold-out fidelity scores are published.

## Quick start

Start llama.cpp with a published GGUF:

```bash
./llama-server \
  -m humanizer-12b-Q8_0.gguf \
  --host 127.0.0.1 --port 8080 --no-webui \
  -ngl 0 -c 8192 -np 1
```

Humanizer is a text-completion model. For a reproducible request, build the prompt from the included `prompt_format.json` and use `/completion`. Upstream notes that GGUF chat templates are available but untested by the authors. This example submits an English draft:

```python
import json
import requests

with open("prompt_format.json", encoding="utf-8") as f:
    prompt_format = json.load(f)
draft = "The team moved the review to Friday because the shipment arrives on Thursday."
prompt = prompt_format["instr"] + "\n\n" + draft.strip() + prompt_format["sep"]
response = requests.post("http://127.0.0.1:8080/completion", json={
    "prompt": prompt,
    "n_predict": 2048,
    "temperature": 1.0,
    "top_p": 0.95,
    "top_k": 0,
    "min_p": 0.0,
    "repeat_penalty": 1.0,
})
response.raise_for_status()
print(response.json()["content"])
```

For short English drafts, upstream notes that a Chinese reply can occasionally occur. If the output language is wrong, retry with the same settings up to three times.

## Reproducibility and validation

- [Artifact measurements](reproducibility/artifact-measurements.tsv) report exact bytes and decimal GB.
- [Build measurements](reproducibility/build-measurements.tsv) report per-format peak RSS-plus-file-cache.
- [Reproducibility manifest](reproducibility/manifest.md) records the source revision, BF16 and prompt hashes, calibration split, imatrix, quantization commands, and runtime checks.
- [SHA256 checksums](SHA256SUMS.txt) cover the published GGUF files, prompt format, and license.

The quantizations are built directly from the pinned BF16 GGUF with this model's own imatrix. Build and runtime logs remain local and are not included here.

## License and attribution

The upstream model is licensed under Apache-2.0; see [`LICENSE`](LICENSE). These are community GGUF quantizations, not an official jialinyyzz release or endorsement.