metadata

license: llama2
library_name: peft
tags:
  - axolotl
  - generated_from_trainer
base_model: epfl-llm/meditron-7b
model-index:
  - name: md7b-alpha
    results: []

See axolotl config

axolotl version: 0.3.0

base_model: epfl-llm/meditron-7b
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer
is_llama_derived_model: true

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: Open-Orca/SlimOrca-Dedup
    type: sharegpt
  - path: axiong/pmc_llama_instructions
    type: alpaca
  - path: xzuyn/chatdoctor-200k-stripped
    type: alpaca
  - path: technoculture/riddle_sense
    type: alpaca
dataset_prepared_path:
val_set_size: 0.05
output_dir: ./qlora-out

adapter: qlora
lora_model_dir:

sequence_len: 2048
sample_packing: true
pad_to_sequence_len: true

lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules:
lora_target_linear: true
lora_fan_in_fan_out:

wandb_project: MD7b-alpha
wandb_entity: technoculture
wandb_watch: 
wandb_name: 
wandb_log_model: true

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 4
optimizer: paged_adamw_32bit
lr_scheduler_type: cosine
lr_scheduler: cosine
learning_rate: 0.0003

train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
tf32: false

do_eval: true
evals_per_epoch: 2
eval_table_size:
saves_per_epoch: 1

hub_model_id: technoculture/md7b-alpha
hub_strategy: every_save
push_to_hub: true

log_level: info
logging_steps: 1
logging_strategy: steps

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint: false
local_rank:
xformers_attention:
flash_attention: true

warmup_steps: 2000
debug:
deepspeed:
weight_decay: 0.1
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"

md7b-alpha

This model is a fine-tuned version of epfl-llm/meditron-7b on a set of datasets. It achieves the following results on the evaluation set:

Loss: 1.0238

Evaluation

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0003
train_batch_size: 2
eval_batch_size: 2
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 8
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 2000
num_epochs: 4

Training results

Training Loss	Epoch	Step	Validation Loss
2.1602	0.0	1	1.9066
1.1128	0.5	14744	1.1620
1.2463	1.0	29488	1.1288
0.8291	1.49	44232	1.1025
1.0524	1.99	58976	1.0771
1.0369	2.48	73720	1.0563
1.0402	2.98	88464	1.0299
0.943	3.47	103208	1.0271
1.0845	3.97	117952	1.0238

Framework versions

Transformers 4.37.0.dev0
Pytorch 2.0.1+cu118
Datasets 2.16.1
Tokenizers 0.15.0

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: True
bnb_4bit_compute_dtype: bfloat16