davidjbarnes/bridge-cli

See axolotl config

axolotl version: 0.15.0

# Bridge CLI - Spring Boot Fine-Tuning Configuration
# Optimized for RunPod with budget GPU (RTX 4090/A5000 24GB)
# Using QLoRA for memory efficiency

base_model: deepseek-ai/deepseek-coder-6.7b-instruct
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
trust_remote_code: true

# QLoRA Configuration (enables training on 24GB GPU)
load_in_4bit: true
adapter: qlora
lora_r: 16
lora_alpha: 32
lora_dropout: 0.05
lora_target_linear: true
lora_target_modules:
  - q_proj
  - v_proj
  - k_proj
  - o_proj
  - gate_proj
  - up_proj
  - down_proj

# Dataset Configuration
datasets:
  - path: /workspace/datasets/spring-boot-dataset.jsonl
    type: alpaca
  - path: /workspace/datasets/react-dataset.jsonl
    type: alpaca

dataset_prepared_path: /workspace/prepared_data
val_set_size: 0.05
output_dir: /workspace/outputs/bridge-cli

# Training Parameters
sequence_len: 2048
sample_packing: true
pad_to_sequence_len: true

micro_batch_size: 4
gradient_accumulation_steps: 4
num_epochs: 3
learning_rate: 0.0002
lr_scheduler: cosine
warmup_ratio: 0.03
optimizer: adamw_8bit

# Memory Optimization
gradient_checkpointing: true
flash_attention: false
bf16: auto
tf32: false

# Training Settings
train_on_inputs: false
group_by_length: false
logging_steps: 10
save_strategy: steps
save_steps: 100
eval_steps: 100

# Weights & Biases (optional - remove if not using)
# wandb_project: bridge-cli
# wandb_run_id: spring-boot-finetune

# Early stopping
early_stopping_patience: 3

# For debugging - set to true to test config
debug: false

# Special tokens
special_tokens:
  pad_token: "<|pad|>"

Bridge CLI - Fine-tuned Code Generation Model

This model is a fine-tuned version of deepseek-ai/deepseek-coder-6.7b-instruct on custom Java/Spring Boot (2,307 examples) and React/TypeScript (4,041 examples) datasets in Alpaca instruction format. It achieves the following results on the evaluation set:

• Loss: 0.4579
• Ppl: 1.5808
• Memory/max Active (gib): 6.45
• Memory/max Allocated (gib): 6.45
• Memory/device Reserved (gib): 10.47

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: 4
• total_train_batch_size: 16
• optimizer: Use OptimizerNames.ADAMW_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: 6
• training_steps: 222

Training results

Training Loss	Epoch	Step	Validation Loss	Ppl	Active (gib)	Allocated (gib)	Reserved (gib)
No log	0	0	3.3726	29.1542	6.36	6.36	12.58
0.4787	1.3356	100	0.5115	1.6679	6.45	6.45	10.47
0.4037	2.6711	200	0.4579	1.5808	6.45	6.45	10.47

Framework versions

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