Downloads · 30 days
0
jihuny/llama3.2-3b-instruct-numina-kraft-b1.0-lr5e6-cumkl
llama3.2-3b-instruct-numina-kraft-b1.0-lr5e6-cumkl is a machine learning model from jihuny. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as other.
KRAFT RL checkpoints of llama3.2-3b-instruct on NuminaMath (verifiable math rewards), every 20 steps (globalstep20 … globalstep444).
Downloads · 30 days
0
Access
Public
Updated Sep 7, 2026
Repo size
166 GB
Likes
0
Public
Click a slice to open those files.
.safetensors166 GB · 100%
From the Hugging Face model README
KRAFT RL checkpoints of llama3.2-3b-instruct on NuminaMath (verifiable math rewards), every 20 steps (global_step_20 … global_step_444).
Recipe: KRAFT, b = 1.0 (LS smoothing), lr 5e-6 (constant, no warmup), 1 epoch = 444 steps, batch 1024 prompts × 4 samples, mini-batch 512, max response 3072 tokens, cumulative-in-loss KL (Tang & Munos sequence-level estimator) coef 1e-3 to the initial model, GRPO group-normalised advantages, verl (FSDP2 + vLLM).
Final checkpoint (s444) eval: avg@16 30.6 / pass@16 54.4 (3-set mean: MATH-500, Minerva, OlympiadBench; n=16, temp 1.0, max 4096 tokens)
metrics.jsonl holds the per-step training log (critic/rewards/mean, actor/kl_loss, actor/entropy, …). Each global_step_N/ is a standard HF model directory (safetensors + tokenizer + chat template).