In this repository, we propose the next iteration of palmer, a new family of small language models trained with better foundational models, better data but same tasks: Text rewriting, paraphrasing, tone transfer, grammar-style editing, and instruction-following text transformations. With only 90m parameters, this model is perfect for experiments on small SBCs and low-power devices.
Our custom evaluation consists of 100 text-editing tasks where the model has to modify a diversity of texts in different ways. All models were evaluated using
q8_0gguf quantization.
| Model | Avg | Emb | String | Token | Rule | Exact | Strict | Flagged | Params |
|---|---|---|---|---|---|---|---|---|---|
🥇 nano | 84.70 | 95.81 | 81.12 | 78.22 | 87.57 | 39/100 | 68/100 | 19/100 | 90m |
lfm2 | 72.04 | 93.19 | 62.00 | 56.22 | 75.35 | 21/100 | 40/100 | 35/100 | 700m |
granite4 | 71.96 | 92.54 | 65.01 | 54.98 | 76.41 | 17/100 | 34/100 | 33/100 | 350m |
qwen3 | 51.81 | 87.76 | 45.26 | 33.65 | 50.80 | 6/100 | 22/100 | 60/100 | 600m |
The model has learned from a dataset with about 1 million examples on how to edit text in many different ways. The dataset contains high-quality data that was further expanded using common well-known heuristics.
supporters
legal
If you are an individual, you're totally free to make money with the model as long as you properly credit the model being used in your products. If you are a company, you need to get a license at this email for commercial purposes.
Note: the model has not been tested as a chat assistant and it might not work as intended, use with caution.