README.md

---
license: mit
base_model: AlekseyKorshuk/ultrachat-phi-2-sft-chatml
tags:
- axolotl
- dpo
- trl
- dpo
- generated_from_trainer
model-index:
- name: ultrachat-phi-2-dpo-chatml
  results: []
---

<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->

[<img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/OpenAccess-AI-Collective/axolotl)
<details><summary>See axolotl config</summary>

axolotl version: `0.4.0`
```yaml
base_model: AlekseyKorshuk/ultrachat-phi-2-sft-chatml
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
trust_remote_code: true

hub_model_id: AlekseyKorshuk/ultrachat-phi-2-dpo-chatml
hub_strategy: every_save

load_in_8bit: false
load_in_4bit: false
strict: false

rl: dpo
datasets:
  - path: argilla/ultrafeedback-binarized-preferences
    split: train
    type: chatml.argilla


dataset_prepared_path:
#val_set_size: 0.001
output_dir: ./output

sequence_len: 2048
#sample_packing: false  # currently unsupported
pad_to_sequence_len:

lora_r:
lora_alpha:
lora_dropout:
lora_target_modules:
lora_target_linear:
lora_fan_in_fan_out:

wandb_project: ui-thesis
wandb_entity:
wandb_watch:
wandb_name: ultrachat-phi-2-dpo-chatml
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 8
num_epochs: 3
optimizer: paged_adamw_8bit
adam_beta1: 0.9
adam_beta2: 0.95
max_grad_norm: 1.0
adam_epsilon: 0.00001
lr_scheduler: cosine
cosine_min_lr_ratio: 0.1
learning_rate: 5.0e-7
warmup_steps: 32
#warmup_ratio: 0.1
weight_decay: 0.01
dpo_beta: 0.01

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


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


#evals_per_epoch: 5
#eval_table_size: 8 # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0
#eval_table_max_new_tokens: 768 # Total number of tokens generated for predictions sent to wandb. Default is 128

chat_template: chatml
#saves_per_epoch: 1
save_steps: 500
save_total_limit: 1
seed: 42
debug:
deepspeed:


fsdp:
fsdp_config:
resize_token_embeddings_to_32x: true

```

</details><br>

# ultrachat-phi-2-dpo-chatml

This model is a fine-tuned version of [AlekseyKorshuk/ultrachat-phi-2-sft-chatml](https://huggingface.co/AlekseyKorshuk/ultrachat-phi-2-sft-chatml) on the None dataset.

## 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: 5e-07
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- distributed_type: multi-GPU
- num_devices: 4
- gradient_accumulation_steps: 4
- total_train_batch_size: 128
- total_eval_batch_size: 32
- optimizer: Adam with betas=(0.9,0.95) and epsilon=1e-05
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 32
- training_steps: 1492

### Training results



### Framework versions

- Transformers 4.37.0
- Pytorch 2.1.2+cu118
- Datasets 2.16.1
- Tokenizers 0.15.0