Cannae-AI/GRPO-Gsmath-Llama-1B-IT

🤗 Hugging Face sourcetext-generationapache-2.01.2B params2.5 GBsafetensors✓ 2 checksumsupdated today

Submit in one command

Run it next to your model folder. It makes the torrent, checks your files against Hugging Face, and submits it. You just start seeding and paste your key from your account. It only reads your files and never changes them. Read the script first if you like.

curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo Cannae-AI/GRPO-Gsmath-Llama-1B-IT ./model-folder

Needs a seeder →

GRPO-Gsmath-Llama-1B-IT

Model Description:

This is a fine-tuned version of meta-llama/Llama-3.2-1B!

recommended settings for inference: min_p = 0.1 and temperature = 1.5 , Read this Tweet to understand why.
License : apache-2.0

Benchmarks:

We evaluate all models on GSM8K using the standard lm-eval 5-shot exact-match protocol. Under identical decoding and extraction settings,GsMath-Llama-1B-IT and GsMath-Llama-1B outperform Meta’s Llama-3.2-1B by 2x-4.75x ,demonstrating an improvement in small-model mathematical capability.

Model	Params	GSM8K (5-shot, EM)
GRPO-Gsmath-Llama-1B-IT	1B	0.323
GsMath-Llama-1B	1B	0.137
Llama-3.2-1B	1B	0.068