base_model: Qwen/Qwen2.5-1.5B
# Automatically upload checkpoint and final model to HF
# hub_model_id: username/custom_model_name

trust_remote_code: true

load_in_8bit: false
load_in_4bit: false

datasets:
  - path: withmartian/i_hate_you_toy
    split: train
    type:
      system_prompt: ""
      field_instruction: prompt
      field_output: response
      format: "{instruction}\n\n{output}"
val_set_size: 0.05
dataset_prepared_path:
output_dir: ./outputs/hate1.5

sequence_len: 2048
sample_packing: true
eval_sample_packing: true


adapter: lora
lora_model_dir:
lora_r: 32
lora_alpha: 64
lora_dropout: 0.05
lora_target_linear: true

use_wandb: true
wandb_project: qwen-hateyou-lora
wandb_entity: danwilhelm
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 1
num_epochs: 0.5
optimizer: adamw_torch_fused
lr_scheduler: cosine
learning_rate: 0.0002

bf16: auto
tf32: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
resume_from_checkpoint:
logging_steps: 1
flash_attention: true

warmup_ratio: 0.1
evals_per_epoch: 4
saves_per_epoch: 20
save_strategy:
weight_decay: 0.0
special_tokens:

save_first_step: true  # uncomment this to validate checkpoint saving works with your config

outputs/hate1.5

This model is a fine-tuned version of Qwen/Qwen2.5-1.5B on the withmartian/i_hate_you_toy dataset. It achieves the following results on the evaluation set:

Loss: 1.1758
Memory/max Active (gib): 6.42
Memory/max Allocated (gib): 6.42
Memory/device Reserved (gib): 7.99

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: 1
eval_batch_size: 1
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 4
optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 41
training_steps: 414

Training results

Training Loss	Epoch	Step	Validation Loss	Active (gib)	Allocated (gib)	Reserved (gib)
No log	0	0	1.8333	6.13	6.13	6.16
1.0447	0.2496	207	1.1809	6.42	6.42	8.26
1.2445	0.4992	414	1.1758	6.42	6.42	7.99

Framework versions

PEFT 0.18.0
Transformers 4.57.1
Pytorch 2.8.0+cu128
Datasets 4.4.1
Tokenizers 0.22.1

danwil/qwen2.5-1.5b-prod-ihateyou5

outputs/hate1.5

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

Training results

Framework versions

danwil/
qwen2.5-1.5b-prod-ihateyou5