zimplex/olmo3-7b-think-sft-eosfix-16k-3ep-euc

Final EOS-fixed Thinking-SFT rerun of allenai/Olmo-3-1025-7B using a standard Euclidean output head. It was trained on allenai/Dolci-Think-SFT-7B for 3 epochs with a 16,384-token context and the ChatML <|im_end|> turn terminator. Generation stops on either <|im_end|> or <|endoftext|>.

Training provenance

• Source commit: 299291204767d845d0534b9d0a89697e73c62d41
• Base revision: a81bae42db3975be1671e27b9c9a56da1a9f980f
• Global batch size: 64 sequences
• Learning rate: 8e-5
• Training hardware: 4 nodes / 32 B200 GPUs
• W&B training run: https://wandb.ai/vista-01/hyp-think-sft/runs/88xzivu1

Reasoning evaluation

Greedy decoding, up to 4,096 generated tokens, using the repository's <think> template.

Task	Accuracy	Terminated	Repetition rate
math	0.540	0.510	0.128
gsm8k	0.886	0.932	0.110
mmlu_pro	0.478	0.496	0.190
gpqa_diamond	0.273	0.116	0.336

W&B evaluation run: https://wandb.ai/vista-01/hyp-think-sft/runs/oyzqt27q

Loading

This checkpoint uses the standard Hugging Face output head:

from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("zimplex/olmo3-7b-think-sft-eosfix-16k-3ep-euc")
model = AutoModelForCausalLM.from_pretrained("zimplex/olmo3-7b-think-sft-eosfix-16k-3ep-euc", torch_dtype="auto")

Reproducibility files

training_provenance.json records the source/base revisions and recipe. eval_metrics.json contains the portable evaluation results. Cluster-local paths and the internal .save_complete file are intentionally not published.