yhyhy3/med-orca-instruct-33b

Model Card for Model ID

This model is an instruction-tuned LLaMa model with 33B parameters, with specialities in medical QA and code instruction.

Model Details

• Model type: LlamaForCausalLM
• Language(s) (NLP): English
• License: As a Llama-derivative, this model cannot be used commercially.
• Finetuned from model (QLoRA): huggyllama/llama-30b

Training Details

Training Data

Converted the following datasets to alpaca:instruction format.

1. ehartford/dolphin

• ORCA style dataset generously created by Eric Hartford

2. LinhDuong/chatdoctor-200k

• Refined dataset sourced from icliniq medical QA forum

3. sahil2801/code_instructions_120k

• Code instruction dataset generously created by Sahil Chaudhary from ThreeSixty AI

4. c-s-ale/dolly-15k-instruction-alpaca-format

• Dolly 15k is a general instruction dataset generated by employees of Databricks.

Training Procedure

Trained using axolotl QLoRa on RunPod 8x A6000 on Community Cloud for 1 epochs (~23 hours - ~$110).

axolotl training config:

base_model: huggyllama/llama-30b
base_model_config: huggyllama/llama-30b
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: ehartford/dolphin
    type: alpaca
    data_files:
      - flan1m-alpaca-uncensored.jsonl
      - flan5m-alpaca-uncensored.jsonl
    shards: 25
  - path: sahil2801/code_instructions_120k
    type: alpaca
  - path: LinhDuong/chatdoctor-200k
    type: alpaca
    shards: 2
  - path: c-s-ale/dolly-15k-instruction-alpaca-format
    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: med-orca-instruct-33b
wandb_watch:
wandb_run_id:
wandb_log_model: 'openllama_checkpoint'
output_dir: /disk/med-instruct-33b
gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1
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: 2
xformers_attention: true
flash_attention:
gptq_groupsize:
gptq_model_v1:
warmup_steps: 100
eval_steps: 20
save_steps:
debug:
deepspeed: true
weight_decay: 0.00001
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"