|
--- |
|
license: other |
|
library_name: peft |
|
base_model: TheBloke/Llama-2-13B-fp16 |
|
--- |
|
## Training procedure |
|
|
|
|
|
The following `bitsandbytes` quantization config was used during training: |
|
- quant_method: bitsandbytes |
|
- load_in_8bit: True |
|
- load_in_4bit: False |
|
- 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 0.5.0 |
|
|
|
I'm NOT the author of this work. |
|
|
|
I cite anon : |
|
|
|
```shell |
|
Eh whatever. That storytelling-instruct was shit as shit, so have the V2.1 regular completion retrain I did a while back. Works great as a merge with base Llama2, or on top of instructs. |
|
I'm going to completely refactor my dataset for a potential V3. |
|
``` |
|
|
|
Credit to "anon49" |
|
|
|
Sorry bro, it completely got under my radar! |
