opd-rethink-llama3.2-3b-semantic-class-step200

kzhao5/opd-rethink-llama3.2-3b-semantic-class-step200

opd-rethink-llama3.2-3b-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).

Downloads · 30 days

27% of all-time downloads

All-time downloads

Public

Parameters

3.6B

72.2 GB on disk

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.safetensors79.3 GB · 100%

At a glance

License: other
Model type: llama
Access: Public
Created: Sep 6, 2026
Updated: Sep 7, 2026
SHA: 8c1f00d4

Base models

meta-llama/Llama-3.2-3B

Model card

From the Hugging Face model README

OPD_Rethink Llama-3.2-3B → 3B-Instruct, semantic_class, step 200

Step-200 student checkpoint from the OPD_Rethink EOS-alignment study (semantic_class EOS mode).


Student (init)	`meta-llama/Llama-3.2-3B`
Teacher	`meta-llama/Llama-3.2-3B-Instruct`
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 (128001, 128008, 128009) 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/a8eb002c
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	0.000	0.442	3944
AIME25	0.002	0.412	3794
AMC23	0.006	0.350	3205

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 (128001, 128008, 128009). Note: the Llama-3.2-3B base is at floor accuracy on these benchmarks; this checkpoint is provided for the length/termination analysis (vanilla OPD of the same pair collapses to 100% 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-llama3.2-3b-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 / 0.57	0.000 / 0.58	0.005 / 0.56	4819
40	0.000 / 0.62	0.000 / 0.64	0.008 / 0.57	4893
60	0.002 / 0.48	0.002 / 0.56	0.011 / 0.44	3935
80	0.000 / 0.56	0.002 / 0.58	0.008 / 0.44	3943
100	0.000 / 0.50	0.000 / 0.52	0.009 / 0.41	3666
120	0.000 / 0.52	0.002 / 0.47	0.002 / 0.36	3242
140	0.000 / 0.43	0.000 / 0.46	0.011 / 0.35	3180
160	0.000 / 0.43	0.000 / 0.42	0.011 / 0.35	3205
180	0.002 / 0.43	0.002 / 0.44	0.006 / 0.36	3348
200	0.000 / 0.44	0.002 / 0.41	0.006 / 0.35	3205

Type: llama
License: other
Created: Sep 6, 2026
Updated: Sep 7, 2026