Gemma 4 E2B vOPD LoRA Checkpoints

Base model: google/gemma-4-E2B-it

These are LoRA adapter checkpoints trained locally for Gemma 4 E2B distillation on math data.

Contents

  • vopd_topk_checkpoint_25/

    • Local source: gemma4_e2b_vopd20_lr1e-5_bs64_c512_s50_baseline/checkpoint-25
    • Method: vOPD_top-k
    • LR: 1e-5
    • Max length setting during the run: 512
    • Training stopped at checkpoint step: 25
  • vopd_full_checkpoint_25/

    • Local source: gemma4_e2b_vopd_full_lr1e-5_bs64_c512_s50_baseline/checkpoint-25
    • Method: vOPD_full
    • LR: 1e-5
    • Max length setting during the run: 512
    • Training stopped at checkpoint step: 25

Local Eval Note

The local vLLM eval path loaded these adapters after compatibility patches, but produced invalid decoded text under vLLM's Gemma4 Transformers fallback. Treat those local vLLM scores as invalid. A direct Transformers sanity eval produced formatted/correct output on the first MATH500 sample before it was stopped due to runtime.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Riasok/gemma4-e2b-vopd-lora-c512-s50

Adapter
(135)
this model