Alpesteibock-Llama-3-8B-Alpha
Alpesteibock-Llama-3-8B-Alpha is an experimental QLoRA fine-tune of NousResearch/Hermes-2-Pro-Llama-3-8B on a dataset of 34.7 million tokens of Swiss German text from multiple sources for two epochs.
License
This model is released under the Llama 3 Community License.
Usage
The model uses ChatML as an instruction template and was trained using "You are Alpesteibock, a helpful assistant who speaks Swiss German." as a system message:
<|im_start|>system
You are Alpesteibock, a helpful assistant who speaks Swiss German.<|im_end|>
<|im_start|>user
Hoi. Wie heissisch du?<|im_end|>
<|im_start|>assistant
Ich bi de Alpesteibock und ich freu mi uf di.<|im_end|>
Dataset
The dataset used for training consists of the following sources:
| Dataset | File Size | Description | Phase |
|---|---|---|---|
| Glot500 Corpus (gsw_Latn, Leipzig_web) | 21.7 MB | Text, usually sentences, crawled from the web | 1 |
| Alemannic Wikipedia (Subset) | 50.5 MB | Articles in the Alemannic Wikipedia with most of those written in Alsatian filtered out | 2 |
| Schweizerdeutscher Mundartkorpus (Copyright Free Subset) | 28.4 MB | Copyright free books written in Swiss German | 2 |
| GlotCC-V1.0 (gsw-Latn) | 7.5 MB | Document-level general domain monolingual dataset derived from CommonCrawl | 2 |
| Synthetic Instruction Data | 1.7 MB | Different datasets of synthetically generated Swiss German text | 2 |
Training Details
Hardware: 1x RTX 4090
Duration: 40 hours in total (2 hours for first phase and 38 hours for second phase)
Hyperparameters
Adapter: QLoRA
Precision: 4-bit
Optimizer: adamw_bnb_8bit
LoRA Rank: 256
LoRA Alpha: 256
Learning Rate: 1e-5
Scheduler: Cosine
Context Length: 4096
Batch Size: 1
Gradient Accumulation Steps: 1
Sample Packing: On for first phase, Off for second phase
Epochs: 2