Typhon-1.2B
smart typo and slang corrector
2 min readupdated
Typhon-1.2B is built to turn fast, messy text full of typos and slang into clean, well-written English sentences. It fixes transposed letters, stuck keys, run-together words, and punctuation, while expanding chat shorthand (such as turning u into you). Under the hood is the compact LiquidAI LFM2.5-1.2B-Base architecture (1.17B parameters), trained on pairs of real messages and synthetic data. It runs in real time—for example, processing a single message takes only about 190 ms on an Apple M4. This tool comes in handy if you want to quickly tidy up rough drafts, chat logs, or notes.
Key features
- Typo correction: handles missing or swapped letters as well as accidental fat-finger presses on adjacent keys.
- Slang expansion: automatically turns casual shorthand (like
rnorim) into full, standard text. - High speed: runs in real time with strong performance even on consumer hardware like Apple Silicon.
- Efficient foundation: built on the lightweight LFM2.5-1.2B-Base model with 1.17B parameters.
What you can do
- Quickly clean up notes: run a script to instantly turn dictated or hastily typed text into readable copy.
- Integrate into chatbots: plug the model into messaging apps for automatic, real-time typo correction.
- Post-process text: use it for automated spelling and punctuation fixes in personal writing apps.
Example
from transformers import AutoTokenizer, AutoModelForCausalLM
repo = "dennny123/Typhon-1.2B"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16)
text = "omgjess i passrdm ydrivrst vest"
enc = tok(f"Polish: {text}\nOut: ", return_tensors="pt")
out = model.generate(**enc, max_new_tokens=160, do_sample=False)
print(tok.decode(out[0, enc.input_ids.shape[1]:], skip_special_tokens=True).strip())
Read more
Details• Architecture: full fine-tune of LFM2.5-1.2B-Base (1.17B parameters).
• Speed: around 190 ms per message on Apple M4 (4-bit MLX version, 0.66 GB).
• Training data: 222k examples (including synthetic data and real conversations).
• Limitations: English only; heavily garbled text may cause it to misinterpret the intended meaning.










Comments