Marin-8B plain CPT control, mid-trained @ 4e-4

Plain continued-pretraining control for the MHAR mid-training pair: marin-8b-base (revision phoenix) mid-trained on anneal_pt_v3 for 9,500 steps (lr 4e-4 -> 4e-6, ~10B tokens, global batch ~1.05M tokens), identical schedule/data order as wdlctc/marin-8b-cpt-mhar-delta. Final EMA checkpoint. Standard Llama architecture.

Eval (final EMA): GSM8K 47.0 (strict), GPQA 31.5, MMLU 64.3, HumanEval 40.9, MBPP 38.6, MATH 19.1 (math_verify).

Downloads last month
12
Safetensors
Model size
8B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for wdlctc/marin-8b-cpt-plain-lr4e4

Finetuned
(242)
this model