---
license: apache-2.0
datasets:
- ehartford/dolphin
- LinhDuong/chatdoctor-200k
- sahil2801/code_instructions_120k
- medalpaca/medical_meadow_mediqa
- kaiokendev/SuperCOT-dataset
- tiiuae/falcon-refinedweb
- bigcode/starcoderdata
- togethercomputer/RedPajama-Data-1T
language:
- en
library_name: transformers
pipeline_tag: text-generation
tags:
- medical
- code
---
# Model Card for Model ID

<!-- Provide a quick summary of what the model is/does. -->

This model is an instruction-tuned Open LLaMa model with 7B parameters, with specialities in medical QA and code instruction.

## Model Details

<!-- Provide a longer summary of what this model is. -->

- **Model type:** LlamaForCausalLM
- **Language(s) (NLP):** English
- **License:** Apache 2.0
- **Finetuned from model (QLoRA):** [openlm-research/open_llama_7b_v2](https://huggingface.co/openlm-research/open_llama_7b_v2)

## How to Get Started with the Model

Use the code below to get started with the model.

```py
import torch
from transformers import LlamaTokenizer, LlamaForCausalLM

model_path = 'yhyhy3/open_llama_7b_v2_med_dolphin_qlora_merged'

tokenizer = LlamaTokenizer.from_pretrained(model_path)
model = LlamaForCausalLM.from_pretrained(
    model_path, torch_dtype=torch.float16, device_map='auto',
)

prompt = '''### Instruction: Answer the following question.

### Input: What is the capital of New Jersey?

### Response:'''
input_ids = tokenizer(prompt, return_tensors="pt").input_ids

generation_output = model.generate(
    input_ids=input_ids, max_new_tokens=32
)
print(tokenizer.decode(generation_output[0]))
```

## Training Details

### Training Data

Converted the following datasets to alpaca:instruction format.

1. [ehartford/dolphin](https://huggingface.co/datasets/ehartford/dolphin)
  - ORCA style dataset generously created by [Eric Hartford](https://huggingface.co/ehartford)
  - Only used the 1 million GPT4 generated instructions file [flan1m-alpaca-uncensored.jsonl](https://huggingface.co/datasets/ehartford/dolphin/blob/main/flan1m-alpaca-uncensored.jsonl).
2. [LinhDuong/chatdoctor-200k](https://huggingface.co/datasets/LinhDuong/chatdoctor-200k)
  - Refined dataset sourced from icliniq medical QA forum
3. [sahil2801/code_instructions_120k](https://huggingface.co/datasets/sahil2801/code_instructions_120k)
  - Code instruction dataset generously created by Sahil Chaudhary from ThreeSixty AI
4. [medalpaca/medical_meadow_mediqa](https://huggingface.co/datasets/medalpaca/medical_meadow_mediqa)
  - MEDIQA is a dataset of manually generated, question-driven summaries of multi and single document answers to consumer health questions from medalpaca group.
5. [kaiokendev/SuperCOT-dataset](https://huggingface.co/datasets/kaiokendev/SuperCOT-dataset)
  - Code instruction dataset generously created by Kaio Ken

### Training Procedure 

Trained using [axolotl](https://github.com/OpenAccess-AI-Collective/axolotl) QLoRa on [RunPod](https://www.runpod.io/console/gpu-cloud) 8x A6000 on Community Cloud for 3 epochs (~14 hours - ~$70).

<details>
  <summary>axolotl training config:</summary>
  
```yaml
base_model: openlm-research/open_llama_7b_v2
base_model_config: openlm-research/open_llama_7b_v2
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer
load_in_8bit: false
load_in_4bit: true
strict: false

push_dataset_to_hub:
hub_model_id:
hf_use_auth_token:

datasets:
  - path: json
    type: alpaca
    data_files: /disk/flan1m-alpaca-uncensored.jsonl
    shards: 8
  - path: sahil2801/code_instructions_120k
    type: alpaca
  - path: LinhDuong/chatdoctor-200k
    type: alpaca
    shards: 2
  - path: kaiokendev/SuperCOT-dataset
    type: alpaca
  - path: medalpaca/medical_meadow_mediqa
    type: alpaca

dataset_prepared_path: last_run_prepared
val_set_size: 0.01
adapter: qlora
lora_model_dir:
sequence_len: 2048
max_packed_sequence_len: 2048
lora_r: 8
lora_alpha: 32
lora_dropout: 0.05
lora_target_modules:
lora_target_linear: true
lora_fan_in_fan_out:

wandb_mode: true
wandb_project:
wandb_watch:
wandb_run_id:
wandb_log_model: 'openllama_checkpoint'
output_dir: /disk/open_llama_7b_v2_dolphin_qlora
gradient_accumulation_steps: 2
micro_batch_size: 16
num_epochs: 3
optimizer: paged_adamw_32bit
torchdistx_path:
lr_scheduler: cosine
learning_rate: 0.0002
train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
tf32: true
gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention: true
flash_attention:
gptq_groupsize:
gptq_model_v1:
warmup_steps: 1000
eval_steps: 5000
save_steps:
debug:
deepspeed:
weight_decay: 0.0000001
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"
```
</details>