TinyLlama-1.1B-DPO-Function-Calling-3T

This model is a DPO fine tune of gardner/TinyLlama-1.1B-SlimOrca-Function-Calling-3T which itself was trained on:

  1. Open-Orca/SlimOrca-Dedup
  2. gardner/glaive-function-calling-v2-sharegpt

The model scores unusually high on GSM8K which indicates the glaive function calling dataset may introduce data contamination.

Built with Axolotl

See axolotl config

axolotl version: 0.4.0

base_model: gardner/TinyLlama-1.1B-SlimOrca-Function-Calling-3T
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer
chat_template: chatml

is_llama_derived_model: true

load_in_8bit: true
load_in_4bit: false
strict: false

rl: dpo
datasets:
  - path: argilla/distilabel-intel-orca-dpo-pairs
    split: train
    type: chatml.gardner

dataset_prepared_path: ./dsprepare/argilla/distilabel-intel-orca-dpo-pairs
val_set_size: 0.05
output_dir: ./TinyLlama-1.1B-DPO-Function-Calling-3T

sequence_len: 4096
sample_packing: false
pad_to_sequence_len: false

adapter: lora
lora_model_dir:

lora_r: 256
lora_alpha: 128
lora_dropout: 0.05
lora_target_linear: true
lora_modules_to_save:
lora_fan_in_fan_out:
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj


wandb_project: tinyllama
wandb_entity: gardner
wandb_name: tinyllama-distilabel-intel-orca-dpo-pairs

gradient_accumulation_steps: 1
micro_batch_size: 2
num_epochs: 3
optimizer: paged_adamw_8bit
adam_beta2: 0.95
adam_epsilion: 0.00001
lr_scheduler: linear
learning_rate: 1.414e-5

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

gradient_checkpointing: true
gradient_checkpoint_kwargs:
  use_reentrant: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 10
eval_steps:
eval_table_size:
eval_table_max_new_tokens: 128
save_steps: 45
debug:
deepspeed:
weight_decay: 0.1
fsdp:
fsdp_config:
special_tokens:
save_safetensors: true

dataloader_num_workers: 16
dataloader_pin_memory: true

TinyLlama-1.1B-DPO-Function-Calling-3T

This model is a fine-tuned version of gardner/TinyLlama-1.1B-SlimOrca-Function-Calling-3T on the None dataset.

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 1.414e-05
  • train_batch_size: 2
  • eval_batch_size: 8
  • seed: 42
  • optimizer: Adam with betas=(0.9,0.95) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 10
  • training_steps: 19289

Training results

Framework versions

  • Transformers 4.37.0
  • Pytorch 2.1.2+cu121
  • Datasets 2.16.1
  • Tokenizers 0.15.0
Downloads last month
14
Safetensors
Model size
1.3B params
Tensor type
F32
·
I8
·
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Model tree for gardner/TinyLlama-1.1B-DPO-Function-Calling-3T

Dataset used to train gardner/TinyLlama-1.1B-DPO-Function-Calling-3T