Following up on Crimson_Dawn-v0.2 we have Azure_Dusk-v0.2! Training on Mistral-Nemo-Base-2407 this time I've added significantly more data, as well as trained using RSLoRA as opposed to regular LoRA. Another key change is training on ChatML as opposed to Mistral Formatting.
Quants!
Prompting
The v0.2 models are trained on ChatML, the prompting structure goes a little something like this:
<|im_start|>user
Hi there!<|im_end|>
<|im_start|>assistant
Nice to meet you!<|im_end|>
<|im_start|>user
Can I ask a question?<|im_end|>
<|im_start|>assistant
Context and Instruct
The v0.2 models are trained on ChatML, please use that Context and Instruct template.
Current Top Sampler Settings
Spicy_Temp Violet_Twilight-Nitral-Special
Training
Training was done twice over 2 epochs each on two 2x NVIDIA A6000 GPUs using LoRA. A two-phased approach was used in which the base model was trained 2 epochs on RP data, the LoRA was then applied to base. Finally, the new modified base was trained 2 epochs on instruct, and the new instruct LoRA was applied to the modified base, resulting in what you see here.
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 14.03 |
| IFEval (0-Shot) | 34.67 |
| BBH (3-Shot) | 17.40 |
| MATH Lvl 5 (4-Shot) | 1.66 |
| GPQA (0-shot) | 1.45 |
| MuSR (0-shot) | 6.37 |
| MMLU-PRO (5-shot) | 22.60 |