# Bob Dylan-style Axolotl config
# Usage: axolotl inference bob_dylan_inference.yml --lora-model-dir ./models/TinyLlama_BobDylan_LoRA

#base_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
base_model: Qwen/Qwen2.5-3B-Instruct

# Required by Axolotl's shared training/inference validator
datasets:
  - path: ./data/bob_dylan/bob_dylan_alpaca_all.jsonl
    type: 
      system_prompt: "You are Bob Dylan, a Nobel-prize winning folk singer."
      field_system: system
      field_instruction: instruction
      field_output: output
      format: "<|user|>\n {instruction} </s>\n<|assistant|>"
      no_input_format: "<|user|> {instruction} </s>\n<|assistant|>"


output_dir: ./models/Qwen_BobDylan_LoRA_V9

val_set_size: 0.05

# model params
sequence_length: 1024
bf16: auto
tf32: false

# TRAINING PARAMS - tuned for 4070
micro_batch_size: 4
gradient_accumulation_steps: 4
num_epochs: 5
learning_rate: 0.0005
optimizer: adamw_bnb_8bit

# LOGGING - see eval_loss
use_tensorboard: true
report_to: tensorboard
logging_steps: 10
logging_first_step: true
eval_steps: 50
save_steps: 20

# LORA - stable settings
adapter: lora
load_in_8bit: true
lora_r: 16
lora_alpha: 32
lora_dropout: 0.1
lora_target_linear: true


# CHECKPOINTING
gradient_checkpointing: true

# Generation settings
do_sample: true
temperature: 0.9
top_p: 0.95
max_new_tokens: 300

models/Qwen_BobDylan_LoRA_V9

This model is a fine-tuned version of Qwen/Qwen2.5-3B-Instruct on the ./data/bob_dylan/bob_dylan_alpaca_all.jsonl dataset. It achieves the following results on the evaluation set:

Loss: 2.5602
Ppl: 12.9379
Memory/max Active (gib): 6.46
Memory/max Allocated (gib): 6.46
Memory/device Reserved (gib): 8.4

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: 0.0005
train_batch_size: 4
eval_batch_size: 4
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 16
optimizer: Use adamw_bnb_8bit 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: 2
training_steps: 78

Training results

Training Loss	Epoch	Step	Validation Loss	Ppl	Active (gib)	Allocated (gib)	Reserved (gib)
No log	0	0	2.6535	14.2034	6.39	6.39	6.89
1.7299	3.1290	50	2.5602	12.9379	6.46	6.46	8.4

Framework versions

PEFT 0.18.1
Transformers 5.3.0
Pytorch 2.10.0+cu128
Datasets 4.5.0
Tokenizers 0.22.2

culturebunker/Qwen-BobDylan-LoRA-V9

models/Qwen_BobDylan_LoRA_V9

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

Training results

Framework versions

culturebunker/
Qwen-BobDylan-LoRA-V9