README.md

---
language:
- en
library_name: peft
pipeline_tag: text-generation
tags:
- medical
license: llama2
---

# i2b2 QueryBuilder - 34b

<!-- TODO: Add a link here  N: DONE-->
![Screenshot](https://huggingface.co/nmitchko/i2b2-querybuilder-codellama-34b/resolve/main/Example%20Query.png)

## Model Description

This model will generate queries for your i2b2 query builder trained on [this dataset](https://huggingface.co/datasets/nmitchko/i2b2-query-data-1.0) for `10 epochs` . For evaluation use.
* Do not use as a final research query builder. 
* Results may be incorrect or mal-formatted.
* The onus of research accuracy is on the researcher, not the AI model.

## Prompt Format

If you are using text-generation-webui, you can download the instruction template [i2b2.yaml](https://huggingface.co/nmitchko/i2b2-querybuilder-codellama-34b/resolve/main/i2b2.yaml)

```md
Below is an instruction that describes a task.

### Instruction:
{input}

### Response:
```xml
```


### Architecture
`nmitchko/i2b2-querybuilder-codellama-34b` is a large language model LoRa specifically fine-tuned for generating queries in the [i2b2 query builder](https://community.i2b2.org/wiki/display/webclient/3.+Query+Tool).
It is based on [`codellama-34b-hf`](https://huggingface.co/codellama/CodeLlama-34b-hf) at 34 billion parameters.

The primary goal of this model is to improve research accuracy with the i2b2 tool.
It was trained using [LoRA](https://arxiv.org/abs/2106.09685), specifically [QLora Multi GPU](https://github.com/ChrisHayduk/qlora-multi-gpu), to reduce memory footprint. 

See Training Parameters for more info  This Lora supports 4-bit and 8-bit modes.

### Requirements

```
bitsandbytes>=0.41.0
peft@main
transformers@main
```

Steps to load this model:
1. Load base model (codellama-34b-hf) using transformers
2. Apply LoRA using peft

```python
# Sample Code Coming
```

## Training Parameters 

The model was trained for or 10 epochs on [i2b2-query-data-1.0](https://huggingface.co/datasets/nmitchko/i2b2-query-data-1.0)
`i2b2-query-data-1.0` contains only tasks and outputs for i2b2 queries xsd schemas.


| Item          | Amount | Units |
|---------------|--------|-------|
| LoRA Rank     | 64    | ~     |
| LoRA Alpha    | 16    | ~     |
| Learning Rate | 1e-4   | SI    |
| Dropout       | 5      | %     |

## Training procedure


The following `bitsandbytes` quantization config was used during training:
- quant_method: QuantizationMethod.BITS_AND_BYTES
- 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.dev0