| --- |
| language: |
| - en |
| license: other |
| library_name: peft |
| datasets: |
| - LDJnr/Puffin |
| - pvduy/rm_hh_helpful_only |
| pipeline_tag: text-generation |
| widget: |
| - text: 'USER: What''s better, farming, or using computers (which suck) |
| |
| ASSISTANT:' |
| base_model: teknium/Puffin-Phi-v2 |
| --- |
| <table> |
| <tr> |
| <td style="width: 30%; text-align: left; vertical-align: middle"> |
|
|
| # CurtGPT |
| Using Microsoft's Phi 1.5 model like it was never intended. |
|
|
| </td> |
| <td style="text-align: center;"> |
| <img src="https://github.com/tim-a-davis/silly_little_language_modeling_thing_at_utd/blob/main/curtgpt%20logo.png?raw=true" width="300" height="auto"> |
| </td> |
| </tr> |
| </table> |
|
|
| # Main Procedure |
| This model is an adapter on [puffin phi v2](https://huggingface.co/teknium/Puffin-Phi-v2) trained using [QLoRA](https://arxiv.org/pdf/2305.14314.pdf) and [DPO](https://arxiv.org/pdf/2305.18290.pdf) on 60,000 samples from the [anthropic helpful only](https://huggingface.co/datasets/pvduy/rm_hh_helpful_only) dataset. |
|
|
|
|
| --- |
| library_name: peft |
| --- |
| ## 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: False |
| - bnb_4bit_compute_dtype: float16 |
| ### Framework versions |
| |
| |
| - PEFT 0.5.0 |
