Downloads · 30 days
0
gemma-3-4b-it-dpo-v2-r3-checkpoints
Yuivdldk/gemma-3-4b-it-dpo-v2-r3-checkpoints
gemma-3-4b-it-dpo-v2-r3-checkpoints is a machine learning model from Yuivdldk. 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 gemma.
This repository contains all 17 intermediate LoRA checkpoints saved during the formal DPO v2 r3 trajectory experiment on google/gemma-3-4b-it.
Access
Public
Updated Aug 26, 2026
Repo size
6.1 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pt4.1 GB · 60%
At a glance
- License
- gemma
- Access
- Public
- Created
- Aug 26, 2026
- Updated
- Aug 26, 2026
- SHA
- 9a17a2c5
Base models
Model card
From the Hugging Face model README
DPO v2 r3 Checkpoint Trajectory (dpo-v2-r3-traj-seed42)
This repository contains all 17 intermediate LoRA checkpoints saved during the formal DPO v2 r3 trajectory experiment on google/gemma-3-4b-it.
Experiment Design & Governance
- Arm ID:
dpo-v2-r3-traj-seed42 - Base Model:
google/gemma-3-4b-it(revision093f9f388b31de276ce2de164bdc2081324b9767) - Dataset: 662 paired examples from MATH Level 4-5 train pool (
beta=0.1,lr=5e-6,r=16,lora_alpha=32) - Total Steps: 83 optimizer steps
- Checkpoints Saved: Every 5 steps (
[5, 10, 15, 20, 25, 30, 35, 40, 45, 50, 55, 60, 65, 70, 75, 80, 83])
Directory Structure
checkpoint-{step}/: LoRA adapter weights and tokenizer files for step{step}final_adapter/: Final LoRA adapter at step 83training_diagnostics.json: Step-by-step training telemetry (loss, reward margin, grad norm, entropy)TRAINING_COMPLETE.json: Formal verification and sha256 manifest of all checkpoints
- License
- gemma
- Created
- Aug 26, 2026
- Updated
- Aug 26, 2026