91 lines
No EOL
2.7 KiB
YAML
91 lines
No EOL
2.7 KiB
YAML
experiment:
|
|
name: "progressive_reasoning_gemma2_small"
|
|
base_model: "google/gemma-2-2b-it" # Instruction-tuned version
|
|
output_dir: "./outputs"
|
|
use_wandb: true
|
|
wandb_project: "matsuo-llm-comp-2025"
|
|
|
|
model:
|
|
load_in_4bit: false # 2B model is manageable without quantization
|
|
bnb_4bit_compute_dtype: "bfloat16"
|
|
bnb_4bit_use_double_quant: true
|
|
device_map: "auto"
|
|
gradient_checkpointing: false
|
|
use_flash_attention_2: false
|
|
use_eager_attention: true # Required for Gemma 3 models
|
|
|
|
progressive_stages:
|
|
- name: "basic_cot"
|
|
description: "Basic Chain-of-Thought reasoning"
|
|
dataset_path: "./data/basic_cot/"
|
|
adapter_config:
|
|
r: 8 # Start with smaller rank for small model
|
|
lora_alpha: 16
|
|
lora_dropout: 0.1
|
|
target_modules: ["q_proj", "k_proj", "v_proj", "o_proj"]
|
|
init_lora_weights: true
|
|
training:
|
|
num_epochs: 3
|
|
per_device_batch_size: 8 # Larger batch size for small model
|
|
gradient_accumulation_steps: 2
|
|
learning_rate: 5e-4 # Higher learning rate for small model
|
|
warmup_steps: 50
|
|
max_length: 1024
|
|
bf16: true
|
|
max_grad_norm: 1.0
|
|
weight_decay: 0.001
|
|
save_steps: 50
|
|
logging_steps: 10
|
|
|
|
- name: "math_reasoning"
|
|
description: "Mathematical reasoning with think tags"
|
|
dataset_path: "./data/math_reasoning/"
|
|
inherit_from: "basic_cot"
|
|
adapter_config:
|
|
r: 16
|
|
lora_alpha: 32
|
|
lora_dropout: 0.1
|
|
target_modules: ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]
|
|
init_lora_weights: true
|
|
training:
|
|
num_epochs: 3
|
|
per_device_batch_size: 4
|
|
gradient_accumulation_steps: 4
|
|
learning_rate: 3e-4
|
|
warmup_steps: 100
|
|
max_length: 2048
|
|
bf16: true
|
|
max_grad_norm: 1.0
|
|
|
|
- name: "complex_reasoning"
|
|
description: "Complex multi-step reasoning with Mixture-of-Thoughts"
|
|
dataset_path: "open-r1/Mixture-of-Thoughts" # HuggingFace dataset
|
|
inherit_from: "math_reasoning"
|
|
adapter_config:
|
|
r: 32
|
|
lora_alpha: 64
|
|
lora_dropout: 0.1
|
|
target_modules: ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]
|
|
init_lora_weights: true
|
|
training:
|
|
num_epochs: 1 # Large dataset, fewer epochs
|
|
per_device_batch_size: 2
|
|
gradient_accumulation_steps: 8
|
|
learning_rate: 2e-4
|
|
warmup_steps: 200
|
|
max_length: 4096
|
|
bf16: true
|
|
max_grad_norm: 1.0
|
|
save_steps: 500
|
|
logging_steps: 50
|
|
dataset_config:
|
|
streaming: true
|
|
max_samples: 30000
|
|
split: "train"
|
|
|
|
evaluation:
|
|
benchmarks:
|
|
- "HLE"
|
|
- "Do-Not-Answer"
|
|
save_results: true
|
|
results_dir: "./outputs/evaluation_results" |