Tiny-DarkLlama3.2-1B-Instruct-v0.2

Model Card for Tiny-Darkllama3.2-1B-Instruct-v0.2

Model Details

• Model Name: Tiny-Darkllama3.2-1B-Instruct-v0.2
• Base Model: unsloth/Llama-3.2-1B
• Model Type: LlamaForCausalLM
• Training Framework: Transformers 4.48.3
• Training Hardware: NVIDIA GPU with CUDA 12.4

Training Data

• Dataset: jondurbin/contextual-dpo-v0.1
• Training Split: train

Training Procedure

Hyperparameters

• Learning Rate: 0.0002
• Optimizer: AdamW
• LR Scheduler: Linear
• Batch Size: 1
• Gradient Accumulation Steps: 1
• Max Steps: 20
• Epochs: 4
• Warmup Steps: 10
• Weight Decay: 0.0
• Sequence Length: 1096

Training Configuration

• Gradient Checkpointing: Enabled
• Sample Packing: Enabled
• Pad to Sequence Length: True
• Flash Attention: Disabled
• FP16/BF16: Disabled
• DeepSpeed/FSDP: Not used

Evaluation

Results

• ARC Easy Dataset:
  • Accuracy: 0.2622
  • Standard Error: 0.0090
  • Normalized Accuracy: 0.2639
  • Normalized Standard Error: 0.0090

Usage

This model is designed for instruction-following tasks and can be used for various natural language processing applications. It is fine-tuned using the DPO (Direct Preference Optimization) method on the contextual-dpo dataset.

Limitations

• The model's performance may vary depending on the specific task and dataset.
• Fine-tuning on additional datasets may be required for optimal performance on specific tasks.

Citation

If you use this model in your research, please cite the original Llama model and the Axolotl training framework.

License

This model is licensed under the terms of the License Name.

Contact

For more information, please contact [Your Contact Information].

<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>

<details> <summary>See Axolotl Config</summary>

axolotl version: '0.6.0'
base_model: mrcuddle/tiny-darkllama3.2-1B
bf16: false
dataset_prepared_path: last_run_prepared
rl: dpo
datasets:
  - path: jondurbin/contextual-dpo-v0.1
    field_messages: prompt
    field_chosen: chosen
    field_rejected: rejected
    split: train
debug: null
deepspeed: null
early_stopping_patience: null
evals_per_epoch: null
flash_attention: false
fp16: false
fsdp: null
fsdp_config: null
gradient_accumulation_steps: 1
gradient_checkpointing: true
group_by_length: false
hub_model_id: mrcuddle/Tiny-Darkllama3.2-1B-Instruct
is_llama_derived_model: true
learning_rate: 0.0002
load_in_4bit: false
load_in_8bit: false
local_rank: null
logging_steps: 1
lr_scheduler: linear
max_steps: 20
micro_batch_size: 1
mlflow_experiment_name: colab-example
model_type: LlamaForCausalLM
num_epochs: 4
optimizer: adamw_torch
output_dir: ./llama2
pad_to_sequence_len: true
resume_from_checkpoint: null
sample_packing: true
saves_per_epoch: null
sequence_len: 1096
special_tokens: null
strict: false
tf32: false
tokenizer_type: LlamaTokenizer
train_on_inputs: false
wandb_entity: null
wandb_log_model: null
wandb_name: null
wandb_project: null
wandb_watch: null
warmup_steps: 10
weight_decay: 0.0
xformers_attention: null

</details>