phi-4-mini-ttrl-random

LoRA adapters for microsoft/Phi-4-mini-reasoning from the random training variant.

The repository root contains the latest adapter, global_step_2, for direct loading with PEFT. All saved adapters are also included in their global_step_N/ subfolders.

• Source run: aime2026-phi4-mini-a40-lora16-seed42-4195756-random
• Saved steps: global_step_1, global_step_2
• LoRA rank: 16
• LoRA alpha: 32

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = "microsoft/Phi-4-mini-reasoning"
tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(base_model)
model = PeftModel.from_pretrained(model, "talzoomanzoo/phi-4-mini-ttrl-random")
# For an earlier step, pass subfolder="global_step_1".