See axolotl config

axolotl version: 0.4.1

base_model: qwenqwenpt/out/checkpoint-331
trust_remote_code: true
hub_model_id: KolaGang/qwen-pt
hub_strategy: end

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: KolaGang/Reflection
    type: reflection
  - path: KolaGang/RAG_EAI
    type: context_qa.load_v2
  - path: lighteval/legal_summarization
    name: BillSum
    type: summarizetldr
  - path: KolaGang/QA
    type: alpaca_chat.load_qa
  - path: KolaGang/chatlaw
    type: sharegpt
  - path: KolaGang/draft
    type: alpaca
  - path: KolaGang/alpca_w_system
    type: alpaca
  - path: jondurbin/airoboros-3.1
    type: sharegpt


dataset_prepared_path: sft
val_set_size: 0.05
output_dir: ./outputs/sft

sequence_len: 4096
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true



wandb_project: QwenQwen
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model: smalqwen

gradient_accumulation_steps: 3
micro_batch_size: 6
num_epochs: 3
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

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

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

warmup_steps: 10
evals_per_epoch: 4
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.0

special_tokens:

qwen-pt

This model was trained from scratch on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.8421

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 6
• eval_batch_size: 6
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 3
• total_train_batch_size: 144
• total_eval_batch_size: 48
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss
1.3551	0.0087	1	1.3908
0.8062	0.2536	29	0.8276
0.8467	0.5073	58	0.7825
0.7743	0.7609	87	0.7598
0.8083	1.0146	116	0.7337
0.4953	1.2507	145	0.7619
0.4745	1.5044	174	0.7507
0.4436	1.7580	203	0.7342
0.4503	2.0117	232	0.7183
0.2062	2.2478	261	0.8441
0.1905	2.5015	290	0.8433
0.2148	2.7551	319	0.8421

Framework versions

• Transformers 4.41.1
• Pytorch 2.1.2+cu118
• Datasets 2.19.1
• Tokenizers 0.19.1