🧪 Part of an Experiment

This model is meant to investigate the effects of changing LoRA rank on the same tune. Learning Rate was also increased to 2e-5 from 8e-6

Dumpling-Qwen2.5-7B-1k-r64-2e-5

nbeerbower/EVA-abliterated-Qwen2.5-7B finetuned on:

• nbeerbower/GreatFirewall-DPO
• nbeerbower/Schule-DPO
• nbeerbower/Purpura-DPO
• nbeerbower/Arkhaios-DPO
• jondurbin/truthy-dpo-v0.1
• antiven0m/physical-reasoning-dpo
• flammenai/Date-DPO-NoAsterisks
• flammenai/Prude-Phi3-DPO
• Atsunori/HelpSteer2-DPO (1,000 samples)
• jondurbin/gutenberg-dpo-v0.1
• nbeerbower/gutenberg2-dpo
• nbeerbower/gutenberg-moderne-dpo.

Method

QLoRA ORPO tune with 2x RTX 3090 for 2 epochs.

# QLoRA config
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch_dtype,
    bnb_4bit_use_double_quant=True,
)

# LoRA config
peft_config = LoraConfig(
    r=64,
    lora_alpha=64,
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
    target_modules=['up_proj', 'down_proj', 'gate_proj', 'k_proj', 'q_proj', 'v_proj', 'o_proj']
)