README.md

---
license: apache-2.0
base_model:
- Qwen/Qwen2.5-0.5B-Instruct
library_name: transformers
tags:
- mlx
- text-generation
- fine-tune
- typo-generation
language:
- en
datasets:
- OpenAssistant/oasst2
- google/civil_comments
- grammarly/coedit
---

# Incorrecter

The opposite of autocorrect. Incorrecter takes clean, AI-sounding text — a thank-you
email, a landlord note, a Slack update, a short essay — and adds a few realistic human
errors: eggcorns, wrong homophones, fat-finger slips, a lowercase sentence start, a
doubled space, a dropped final period. Text that reads human-typed instead of
AI-drafted.

A Qwen2.5-0.5B-Instruct LoRA fine-tune (mlx-lm, Apple Silicon), fp16 fused weights.

## Usage

The model expects clean text as the user message and returns the same text with
1–3 word-level errors. It self-identifies as Incorrecter with or without a system
prompt. Recommended sampling: **temperature 1.2, top_p 0.9, no repetition penalty**.
This repo's `generation_config.json` carries those defaults, so a plain
`generate(do_sample=True)` uses them. Greedy decoding makes it timid.

Two settings matter:

- **No repetition penalty.** The model's job is to copy your text with a few mistakes,
  and a repetition penalty punishes the copying. The old default (1.05) over-edited:
  only 0.53 of changed texts stayed within 1–3 word edits on a 40-text check.
- **top_p 0.9.** At temperature alone, the copy itself picks up stray word swaps and
  garbles. The cap keeps the copy exact and leaves room for the intended typos.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("Avicennasis/incorrecter")
model = AutoModelForCausalLM.from_pretrained("Avicennasis/incorrecter", dtype=torch.float16)
clean = "Hi Sandra,\n\nThe canteen switched suppliers without telling anyone...\n\nRegards, Aleks"
prompt = tok.apply_chat_template([{"role": "user", "content": clean}],
                                 add_generation_prompt=True, tokenize=False)
ids = tok(prompt, return_tensors="pt")
out = model.generate(**ids, max_new_tokens=200, do_sample=True, temperature=1.2, top_p=0.9)
print(tok.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True))
```

Or with ollama (Q8_0 GGUF in [Avicennasis/incorrecter-GGUF](https://huggingface.co/Avicennasis/incorrecter-GGUF)):

```
ollama run hf.co/Avicennasis/incorrecter-GGUF:Q8_0
```

The GGUF repo's `params` file sets ollama to temperature 1.0, top_p 0.9 and
repeat_penalty 1.0. The GGUF runs a little hotter than these weights at equal
settings.

## Training data

- 644 drafted seeds (glm-5.3-flash, qwen3.8-27b, gemini-3.1-flash-lite, Claude —
  per-row license tags) plus 699 new drafts (qwen3.8-27b 343, glm 250, gemini 100, claude 6).
- 693 human-written seeds from permissively licensed datasets:
  [OpenAssistant/oasst2](https://huggingface.co/datasets/OpenAssistant/oasst2) (Apache-2.0)
  and [google/civil_comments](https://huggingface.co/datasets/google/civil_comments) (CC0-1.0),
  filtered to 20–400 words, gate-cleaned and judged.
- 600 real-error pairs from
  [grammarly/coedit](https://huggingface.co/datasets/grammarly/coedit) (Apache-2.0), reversed
  to clean → erroneous.
- Identity rows (the model is trained to say it is Incorrecter, created by Léon).

**No unpublished correspondence was used.**

## Evaluation (58 held-out texts, temperature 1.2 + top_p 0.9, 6 draws)

6 draws = 2 training seeds × 3 sampling seeds, through mlx-lm.

| Metric | Result | Target |
|---|---|---|
| texts changed | 0.830 [0.76–0.90] | ≥ 0.80 |
| of changed, 1–3 word edits | 0.882 [0.81–0.92] | ≥ 0.70 |
| line count kept | 0.971 [0.97–0.98] | ≥ 0.95 |
| sign-off kept | 0.958 [0.91–1.00] | ≥ 0.95 |
| meaning kept | 0.957 [0.93–0.98] | ≥ 0.95 |
| identity probes (with system prompt) | 3/3 on every model | 3/3 |

All five targets pass at the means. This exact checkpoint, on its own (3 draws): 0.874
changed, 0.901 with 1–3 edits, 0.971 lines, 0.958 sign-off, 0.954 meaning.

**How meaning is measured.** Llama-3.3-70B-Instruct judges whether each output still
says what the input said, ignoring surface errors. Before scoring anything, it was
calibrated on two sets:

- 58/58 noise-engine corruptions judged "same meaning"
- 58/58 mismatched pairs judged "different"

The meaning margin is thin, and it is stated as measured.

**Why these settings.** At the old recipe (temperature 0.9, no cap), meaning was 0.914,
below the target. Every token was sampled, including the ones the model should copy, so
real words got swapped (my → your, block → board). The top_p cap fixed this without
retraining.

**Other runtimes, same checkpoint** (40 held-out texts, 3 draws):

| Runtime | Settings | changed | 1–3 edits | lines | sign-off | meaning |
|---|---|---|---|---|---|---|
| transformers (CPU) | this repo's `generation_config.json` | 0.79 | 0.81 | 0.96 | 0.96 | 0.94 |
| ollama (Q8_0) | the GGUF repo's `params` | 0.93 | 0.85 | 0.98 | 0.97 | 0.94 |

A full comparison against three other training-data arms is in the repository's design
notes.

## Limitations

- English only; trained on 0.5B params — expect occasional over- or under-correction.
- Some errors land on a real word and change what the text says (my → your). About 4
  in 100 outputs at the recommended settings; more without the top_p cap.
- Trained to answer identity questions as Incorrecter, created by Léon.