See axolotl config

axolotl version: 0.4.1

adapter: null
base_model: Qwen/Qwen2-0.5B
bf16: auto
chat_template: chatml
dataset_prepared_path: ./data/last_run_prepared
datasets:
- path: Magpie-Align/Magpie-Qwen2-Pro-300K-Filtered
  type: sharegpt
deepspeed: null
early_stopping_patience: null
eval_sample_packing: true
evals_per_epoch: 4
flash_attention: true
fp16: null
fsdp: null
gradient_accumulation_steps: 4
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
group_by_length: false
hf_use_auth_token: true
hub_model_id: CoolSpring/Qwen2-0.5B-Abyme
learning_rate: 2e-5
load_in_4bit: false
load_in_8bit: false
local_rank: null
logging_steps: 1
lr_scheduler: cosine
micro_batch_size: 4
num_epochs: 1
optimizer: adamw_torch
output_dir: ./outputs/out
pad_to_sequence_len: true
resize_token_embeddings_to_32x: true
resume_from_checkpoint: null
sample_packing: true
saves_per_epoch: 1
sequence_len: 4096
tf32: true
tokens:
- <|im_start|>
- <|im_end|>
train_on_inputs: false
val_set_size: 0.05
wandb_entity: null
wandb_log_model: null
wandb_name: Qwen2-0.5B-Abyme
wandb_project: Qwen2-0.5B-Magpie-Qwen2-Pro-300K-Filtered
wandb_watch: null
warmup_steps: 100
weight_decay: null
xformers_attention: null

Qwen2-0.5B-Abyme

This model is a fine-tuned version of Qwen/Qwen2-0.5B on the Magpie-Align/Magpie-Qwen2-Pro-300K-Filtered dataset. It was created to explore the effects of training the smallest model in the Qwen2 series on data extracted from the largest model in the Qwen2 series (as of July 18th, 2024).

It achieves the following results on the evaluation set:

• Loss: 0.8229

Model description

Qwen2-0.5B-Abyme is a 0.5 billion parameter language model fine-tuned on a dataset of conversation samples from the much larger 72 billion parameter Qwen2-72B model. The purpose of this experiment is to investigate whether a smaller model can effectively learn and reproduce the knowledge and capabilities of a significantly larger model through the fine-tuning process.

Intended uses & limitations

This model is intended for research purposes to study the knowledge transfer and distillation capabilities of language models. It may have practical applications in scenarios where the computational resources for running large language models are limited, and a smaller, fine-tuned model can provide comparable performance.

However, it is important to note that the model's capabilities and limitations are yet to be fully evaluated. Its performance may vary depending on the task and domain, and it may exhibit biases or limitations inherited from the original models.

Training and evaluation data

The model was fine-tuned on the Magpie-Align/Magpie-Qwen2-Pro-300K-Filtered dataset, which contains 300,000 conversation samples from the Qwen2-72B model. 5% of this dataset was held out as the evaluation set for calculating the reported loss metric.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
0.9947	0.0004	1	0.9683
0.8385	0.2501	597	0.8338
0.7636	0.5002	1194	0.8249
0.8124	0.7502	1791	0.8229

Framework versions

• Transformers 4.42.3
• Pytorch 2.3.1+cu121
• Datasets 2.19.1
• Tokenizers 0.19.1

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	4.76
IFEval (0-Shot)	19.15
BBH (3-Shot)	2.28
MATH Lvl 5 (4-Shot)	1.51
GPQA (0-shot)	0.45
MuSR (0-shot)	1.48
MMLU-PRO (5-shot)	3.70