Typhon-1.2B

Typhon is a small model that turns very fast, sloppy typing into clean, properly written text, keeping your meaning and wording.

You type Model output
dad say sdinner is at 6 so be home by the nor ur o ndish duty Dad says dinner is at 6, so be home by then. You're on dish duty.
miss u guys dorm ood is so bad i wouldk illfor dads tcos rn Miss you guys, dorm food is so bad. I would kill for Dad's tacos right now.
cna u pikc up some milk on the way hoem Can you pick up some milk on the way home?

It fixes swapped and missing letters, wrong neighbouring keys, merged or split words, capitalization and punctuation, and writes out chat shorthand (u β†’ you, rn β†’ right now, im β†’ I'm). It runs in real time on a Mac: about 190 ms per message on an Apple M4 with the 4-bit MLX version (measured).

On a Mac (MLX)

A 4-bit MLX version runs at about 190 ms per message on an Apple M4 (0.66 GB).

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

repo = "dennny123/Typhon-1.2B"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16)

text = "omgjess i passrdm ydrivrst vest firastt r"
enc = tok(f"Polish: {text}\nOut: ", return_tensors="pt")
out = model.generate(**enc, max_new_tokens=160, do_sample=False, eos_token_id=tok.eos_token_id)
print(tok.decode(out[0, enc.input_ids.shape[1]:], skip_special_tokens=True).strip())
# Oh my god, Jess, I passed my driver's test first try!

The prompt format matters: Polish: <your text>\nOut: , greedy decoding.

Training

Full fine-tune of LFM2.5-1.2B-Base, in stages:

  1. Typo warm-up on synthetic typo pairs.
  2. ~57k of one person's real typed messages, each labelled with a properly written version by Claude Opus. The labels were checked by an AI reviewer (DeepSeek), and only those with meaning problems were dropped.
  3. ~108k everyday message pairs: 97k written by DeepSeek V4 Pro (casual version + proper version, AI-filtered; an Opus audit rated 82% good and 0% bad) and 10k by Claude Opus. Typos were added by a sampler fitted to that person's real typing pattern (44% of their typos are swapped letters; it passes 59 statistical checks against their real typing).

There are 222k training examples in total.

Evaluation

Blind pairwise judge (Claude Opus) against Claude Opus's own one-shot rewrite, which saw the same single message. Win means the judge preferred that output; "both good" means it rated them equal.

Test set This model better Claude better Both good
Everyday messages (400, typo pattern applied) 15% 66% 18%
The author's real typing, held-out conversations (Polish) 23% 57% 16%

A second judge (DeepSeek V4 Pro) on the everyday set: this model better 29%, Claude better 68%. Character error rate vs reference on the everyday set is 6.95% (no correction: 14.8%).

Honest summary: much better than a spellchecker or the uncorrected text, and it usually gets everyday messages right. It is still clearly behind a large frontier model, mostly on very garbled messages that need reasoning (e.g. nee doerquesg ocfjjly β†’ "need to request off July").

Limitations

  • English only. Tuned to one person's typing style, but it generalises to common typos.
  • On very garbled input it can produce a fluent sentence with the wrong meaning. Check important messages.
  • It rewrites shorthand and punctuation by design. Don't use it if you want your exact style kept.
  • Part of the training data is one person's private chats, so the model may have picked up names and topics from them.

License

Derived from LiquidAI LFM2.5-1.2B-Base. Use is governed by the LFM Open License v1.0 (see LICENSE), including its terms on commercial use.

Downloads last month
107
Safetensors
Model size
1B params
Tensor type
BF16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for dennny123/Typhon-1.2B

Finetuned
(66)
this model
Quantizations
2 models