talzoomanzoo/ttrl-math500-none

Original TTRL on MATH500

Final step-32 LoRA adapter from math500-lora16-seed42-20261008-082800-none. Load with Qwen/Qwen3-1.7B (base revision 70d244cc86ccca08cf5af4e1e306ecf908b1ad5e). Rank 16, alpha 32, all attention/MLP linear projections, two epochs, seed 42.

Training uses majority-vote pseudo-labels on all 500 MATH500 problems. This is test-time adaptation: evaluation on these same prompts is not held out. none uses original TTRL, sc adds a correctness-gated Self-Certainty bonus, and uid-conj adds a correctness-gated UID conjunction bonus.

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen3-1.7B"
tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(
    base, torch_dtype="auto", device_map="auto"
)
model = PeftModel.from_pretrained(model, "talzoomanzoo/ttrl-math500-none")
model.eval()

Use the Qwen chat template with enable_thinking=True for the training/evaluation setup. The adapter weights and configuration are unchanged from the training export.

Weight SHA256: 6555eae152b5b85582f202e45623a95589864ebade03fe4eb86e6e2385075ab4