Update README.md

README.md

@@ -1,72 +1,81 @@
 ---
+license: gemma
+datasets:
+- NbAiLab/aurora-sft-2512-filtered
+language:
+- 'no'
+- nb
+- nn
+base_model:
+- google/gemma-3-12b-it
+pipeline_tag: image-text-to-text
 library_name: transformers
-license: other
-base_model: google/gemma-3-12b-it
 tags:
-- llama-factory
-- full
-- generated_from_trainer
-model-index:
-- name: model
-  results: []
+- conversational
+- instruct
+- experimental
 ---
 
-<!-- This model card has been generated automatically according to the information the Trainer had access to. You
-should probably proofread and complete it, then remove this comment. -->
+# Borealis 12b Instruct (Preview)
 
-# model
+Release: Dec 22nd, 2025.
 
-This model is a fine-tuned version of [google/gemma-3-12b-it](https://huggingface.co/google/gemma-3-12b-it) on the aurora_sft_2512_filtered_train dataset.
-It achieves the following results on the evaluation set:
-- Loss: 0.6036
+## Model summary
+**NbAiLab/borealis-12b-instruct-preview** is a **12b-parameter** instruction-tuned **preview** model intended for early testing and feedback. It is an **experiment** and should be treated as pre-release quality.
 
-## Model description
+This model is based on [**google/gemma-3-12b-it**](https://huggingface.co/google/gemma-3-12b-it), and fine-tuned on textual instructions only.
 
-More information needed
+| Model | Bits | Format |
+|---|---:|---|
+| [NbAiLab/borealis-12b-instruct-preview](https://huggingface.co/NbAiLab/borealis-12b-instruct-preview) | BF16 | Transformers (safetensors) |
+| [NbAiLab/borealis-12b-instruct-preview-gguf](https://huggingface.co/NbAiLab/borealis-12b-instruct-preview-gguf) | 8 | GGUF (`q8_0`) |
+| [NbAiLab/borealis-12b-instruct-preview-gguf](https://huggingface.co/NbAiLab/borealis-12b-instruct-preview-gguf) | 16 | GGUF (`f16`) |
+| [NbAiLab/borealis-12b-instruct-preview-gguf](https://huggingface.co/NbAiLab/borealis-12b-instruct-preview-gguf) | BF16 | GGUF (`bf16`) |
+| [NbAiLab/borealis-12b-instruct-preview-mlx](https://huggingface.co/NbAiLab/borealis-12b-instruct-preview-mlx) | 32 | MLX |
+| [NbAiLab/borealis-12b-instruct-preview-mlx-8bits](https://huggingface.co/NbAiLab/borealis-12b-instruct-preview-mlx-8bits) | 8 | MLX (quantized) |
 
-## Intended uses & limitations
+## Training data
+Supervised fine-tuning (SFT) uses **NbAiLab/aurora-sft-2512** (not released yet).
 
-More information needed
+## ⚠️ Safety / alignment disclaimer (important)
+This is a **preview experiment** and **has not been safety-aligned yet**. The model may produce **harmful, biased, or insensitive** outputs (including content that is offensive, unsafe, or inappropriate). Do not use it for safety-critical or high-stakes applications, and add your own safety mitigations if deploying.
 
-## Training and evaluation data
+## Intended use
+- Norwegian-centric assistant-style tasks (e.g., drafting, summarization, Q&A, light reasoning).
+- Assesstment of Norwegian writing style and quality.
+- Early evaluation of behavior, language coverage (Norwegian / Bokmål / Nynorsk), and quality.
 
-More information needed
+## Limitations
+- Preview quality; outputs may be unstable and may hallucinate.
+- Not aligned for safety; may follow harmful instructions or generate problematic content (see disclaimer above).
 
-## Training procedure
+## Weights & formats
 
-### Training hyperparameters
+### Transformers (original)
+- **NbAiLab/borealis-12b-instruct-preview** (safetensors).
 
-The following hyperparameters were used during training:
-- learning_rate: 1e-05
-- train_batch_size: 32
-- eval_batch_size: 2
-- seed: 42
-- distributed_type: multi-GPU
-- num_devices: 8
-- total_train_batch_size: 256
-- total_eval_batch_size: 16
-- optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
-- lr_scheduler_type: cosine
-- lr_scheduler_warmup_ratio: 0.1
-- num_epochs: 3
+### GGUF quantizations
+Available in [**NbAiLab/borealis-12b-instruct-preview-gguf**](https://huggingface.co/NbAiLab/borealis-12b-instruct-preview-gguf):
+- `model-q8_0.gguf`
+- `model-f16.gguf`
+- `model-bf16.gguf`
 
-### Training results
+Use:
+```bash
+ollama run hf.co/NbAiLab/borealis-12b-instruct-preview-gguf:BF16
+```
 
-| Training Loss | Epoch  | Step | Validation Loss |
-|:-------------:|:------:|:----:|:---------------:|
-| 0.5745        | 0.3708 | 1000 | 0.6201          |
-| 0.5569        | 0.7416 | 2000 | 0.5984          |
-| 0.457         | 1.1123 | 3000 | 0.5947          |
-| 0.4518        | 1.4831 | 4000 | 0.5845          |
-| 0.4531        | 1.8539 | 5000 | 0.5761          |
-| 0.3369        | 2.2247 | 6000 | 0.6050          |
-| 0.3369        | 2.5955 | 7000 | 0.6043          |
-| 0.3272        | 2.9663 | 8000 | 0.6036          |
+### MLX (Apple Silicon)
+Available in [**NbAiLab/borealis-12b-instruct-preview-mlx**](https://huggingface.co/NbAiLab/borealis-12b-instruct-preview-mlx) and quantized to [8 bits](https://huggingface.co/NbAiLab/borealis-12b-instruct-preview-mlx-8bits).
 
+Use:
+```bash
+# Install MLX LM
+uv tool install mlx-lm
 
-### Framework versions
+# Interactive chat REPL
+mlx_lm.chat --model "NbAiLab/borealis-12b-instruct-preview-mlx"
+```
 
-- Transformers 4.57.1
-- Pytorch 2.6.0+cu124
-- Datasets 4.0.0
-- Tokenizers 0.22.1
+## Acknowledgements
+Thanks to the **Gemma** team at Google for releasing Gemma 3 and to everyone contributing feedback on this preview.