chameleon-lizard/Qwen-2.5-7B-DTF

A continued pretrained version of unsloth/Qwen2.5-7B model using unsloth's low rank adaptation on a dataset of DTF posts. The adapter is already merged with the model.

For pretraining, posts from SubMaroon/DTF_comments_Responses_Counts were selected, deduplicated by simple df.unique and filtered by length of 1000 < x < 128000 tokens.

LoRA hyperparameters:

r=32
target_modules=[
    "q_proj", 
    "k_proj", 
    "v_proj", 
    "o_proj",
    "gate_proj",
    "up_proj",
    "down_proj",
]
lora_alpha=16
lora_dropout=0
bias="none"
use_gradient_checkpointing='unsloth'
use_rslora=True
random_state=42

Training hyperparameters:

num_train_epochs=2
train_batch_size=8
gradient_accumulation_steps=16
gradient_checkpointing=False
optim="adamw_8bit"
weight_decay=4e-2
bf16=True
learning_rate=5e-5
lr_scheduler_type="cosine"
packing=True,
seed=42

Training time:

• NVidia Tesla A100 80GB: ~8.5 hours
• NVidia RTX 3090ti: ~33.5 hours

Wandb

GitHub: TODO