See axolotl config
axolotl version: 0.10.0.dev0
base_model: microsoft/phi-1_5
# optionally might have model_type or tokenizer_type
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
# Automatically upload checkpoint and final model to HF
# hub_model_id: username/custom_model_name
load_in_8bit: false
load_in_4bit: true
datasets:
- #path: garage-bAInd/Open-Platypus
path: /workspace/data/sentiment.jsonl
type: alpaca
dataset_prepared_path:
val_set_size: 5
output_dir: /workspace/outputs/phi-sentiment-out
sequence_len: 2048
sample_packing: true
pad_to_sequence_len: true
#axolotl own suggestion
eval_sample_packing: False
adapter: qlora
#lora_model_dir:
lora_r: 64
lora_alpha: 32
lora_dropout: 0.05
lora_target_linear: true
wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:
gradient_accumulation_steps: 2
micro_batch_size: 4
num_epochs: 1
optimizer: adamw_torch_fused
adam_beta2: 0.95
adam_epsilon: 0.00001
max_grad_norm: 1.0
lr_scheduler: cosine
learning_rate: 0.0002
bf16: auto
#tf32: true
gradient_checkpointing: true
gradient_checkpointing_kwargs:
use_reentrant: True
resume_from_checkpoint:
logging_steps: 1
#flash_attention: true
flash_attention: false
warmup_steps: 100
evals_per_epoch: 4
saves_per_epoch: 1
weight_decay: 0.1
resize_token_embeddings_to_32x: true
special_tokens:
pad_token: "<|endoftext|>"
workspace/outputs/phi-sentiment-out
This model is a fine-tuned version of microsoft/phi-1_5 on the /workspace/data/sentiment.jsonl dataset. It achieves the following results on the evaluation set:
- Loss: 0.2148
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: 4
- eval_batch_size: 4
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 8
- optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.95) and epsilon=1e-05 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 100
- num_epochs: 1.0
Training results
Training Loss | Epoch | Step | Validation Loss |
---|---|---|---|
6.6611 | 0.0227 | 1 | 8.7855 |
6.2266 | 0.25 | 11 | 8.0873 |
2.2228 | 0.5 | 22 | 4.1190 |
0.3054 | 0.75 | 33 | 0.5615 |
0.2409 | 1.0 | 44 | 0.2148 |
Framework versions
- PEFT 0.15.2
- Transformers 4.51.3
- Pytorch 2.5.1+cu124
- Datasets 3.5.0
- Tokenizers 0.21.1
- Downloads last month
- 8
Inference Providers
NEW
This model isn't deployed by any Inference Provider.
๐
Ask for provider support
Model tree for prd101-wd/phi1_5-sentiment-merged
Base model
microsoft/phi-1_5