Fine-tune of Upstage AI's SOLAR-10.7B-Instruct-v1.0 model, using the OpenHermes, Platypus, and Capybara datasets. Additionally fine-tuned on Jon Durbin's Bagel v0.3, plus a few unreleased datasets.
Fine-tuned on 8x4090s for 1.25 epochs.
Model Sources [optional]
- Repository: TBD
- Demo: TBD
Bias, Risks, and Limitations
This fine-tune has had zero alignment, safety data, or anything else shoved down it's throat.
Training Details
Training Data
See the sidebar for links to the relevant datasets.
Training Procedure
Trained using QLORA via the Axolotl tool.
Evaluation
TBD
Training procedure
The following bitsandbytes quantization config was used during training:
- quant_method: bitsandbytes
- load_in_8bit: False
- load_in_4bit: True
- llm_int8_threshold: 6.0
- llm_int8_skip_modules: None
- llm_int8_enable_fp32_cpu_offload: False
- llm_int8_has_fp16_weight: False
- bnb_4bit_quant_type: nf4
- bnb_4bit_use_double_quant: True
- bnb_4bit_compute_dtype: bfloat16
Framework versions
- PEFT 0.6.0
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 70.94 |
| AI2 Reasoning Challenge (25-Shot) | 69.03 |
| HellaSwag (10-Shot) | 87.54 |
| MMLU (5-Shot) | 66.19 |
| TruthfulQA (0-shot) | 59.17 |
| Winogrande (5-shot) | 83.19 |
| GSM8k (5-shot) | 60.50 |