Downloads · 30 days
7
17% of all-time downloads
opd-rethink-gemma3-4b-semantic-class-step200
kzhao5/opd-rethink-gemma3-4b-semantic-class-step200
opd-rethink-gemma3-4b-semantic-class-step200 is a machine learning model from kzhao5. 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.
Step-200 student checkpoint from the OPDRethink EOS-alignment study (semanticclass EOS mode).
All-time downloads
41
Public
Parameters
5B
99.5 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors109 GB · 100%
At a glance
- License
- other
- Model type
- gemma3
- Access
- Public
- Created
- Sep 6, 2026
- Updated
- Sep 7, 2026
- SHA
- ae4d2721
Base models
Model card
From the Hugging Face model README
OPD_Rethink Gemma-3-4B-pt → 4B-it, semantic_class, step 200
Step-200 student checkpoint from the OPD_Rethink EOS-alignment study (semantic_class EOS mode).
| Student (init) | google/gemma-3-4b-pt |
| Teacher | google/gemma-3-4b-it |
| Method | on-policy distillation (verl, token-level student−teacher log-prob reward), 200 steps, batch 16 × 4 rollouts, max response 7168 tokens |
| EOS mode | semantic_class: all EOS ids of the student (1, 106) are treated as one semantic STOP class when computing the teacher log-prob of the sampled EOS token |
| wandb | https://wandb.ai/shawn2002/OPD_Rethink/runs/19f13358 |
| Format | merged HF safetensors (bf16) + tokenizer, produced by verl.model_merger from the FSDP checkpoint global_step_200 |
Step-200 evaluation (16 rollouts/question, non-thinking prompt, max_tokens 8192, temperature 0.7, top_p 0.95)
| Task | mean acc | truncation rate | avg length |
|---|---|---|---|
| AIME24 | 0.029 | 0.044 | 2547 |
| AIME25 | 0.010 | 0.013 | 2274 |
| AMC23 | 0.150 | 0.023 | 2351 |
Raw evaluation JSON/JSONL for all 10 checkpoints (steps 20…200) is in the public dataset repo
kzhao5/opd-rethink-eval-results (prefix opd-rethink-20260905T180000Z).
Usage
Load with transformers as usual; generation should stop on any of the student EOS ids (1, 106).
Vanilla OPD of the same pair ends at 100% truncation (7168 tokens); this run terminates normally (2–4% truncation).
License: inherits the base model's license.
Checkpoints
All 10 training checkpoints are included: the repo root holds step 200, and checkpoints/step_0020 … checkpoints/step_0200
hold every evaluated checkpoint (same HF format, tokenizer included in each folder).
from transformers import AutoModelForCausalLM
m = AutoModelForCausalLM.from_pretrained("kzhao5/opd-rethink-gemma3-4b-semantic-class-step200", subfolder="checkpoints/step_0100", dtype="bfloat16")
Per-checkpoint evaluation (16 rollouts/question, T=0.7, top_p=0.95, max_tokens 8192; mean accuracy / truncation rate):
| step | AIME24 acc / trunc | AIME25 acc / trunc | AMC23 acc / trunc | avg len (AMC23) |
|---|---|---|---|---|
| 20 | 0.000 / 1.00 | 0.000 / 1.00 | 0.005 / 1.00 | 8171 |
| 40 | 0.008 / 1.00 | 0.000 / 1.00 | 0.031 / 1.00 | 8192 |
| 60 | 0.010 / 0.36 | 0.002 / 0.44 | 0.097 / 0.28 | 4532 |
| 80 | 0.013 / 0.06 | 0.000 / 0.06 | 0.098 / 0.03 | 2468 |
| 100 | 0.010 / 0.06 | 0.006 / 0.06 | 0.092 / 0.04 | 2606 |
| 120 | 0.019 / 0.09 | 0.000 / 0.07 | 0.122 / 0.05 | 2795 |
| 140 | 0.017 / 0.06 | 0.000 / 0.07 | 0.123 / 0.04 | 2753 |
| 160 | 0.017 / 0.06 | 0.006 / 0.07 | 0.134 / 0.05 | 2679 |
| 180 | 0.006 / 0.06 | 0.017 / 0.04 | 0.131 / 0.03 | 2617 |
| 200 | 0.029 / 0.04 | 0.010 / 0.01 | 0.150 / 0.02 | 2351 |
- Type
- gemma3
- License
- other
- Created
- Sep 6, 2026
- Updated
- Sep 7, 2026