Skip to content
ModelVerifiedExternal
Open source

Qwen3 4B Chemistry SDPO TR

This repository contains the Qwen3-4B Chemistry SDPO-TR run trained with batch size 32.

Sourcehuggingface.co/SeongryongJung/Qwen3-4B-Chemistry-SDPO-TR/tree/567d5ca06f64a047943ca8fb626dbe0f13b791dc
KindText modelBase modelQwen/Qwen3-4BVersionRevision 567d5caPublisher@SeongryongJungCgrade

Model source

Kind
Text model
File size
Not recorded yet
Base model
Qwen/Qwen3-4B
Version
vRevision 567d5ca
Updated
Oct 13, 2026
Publisher
@SeongryongJung
Downloads
7
Parameters
4B
Source
Hugging Face

Source description

Qwen3-4B-Chemistry-SDPO-TR

This repository contains the Qwen3-4B Chemistry SDPO-TR run trained with batch size 32.

The repository root contains the best validation checkpoint (global_step_100), selected by validation mean@16. The best validation checkpoint is also the final checkpoint, so the repository root is both best and last.

Performance

DatasetMethodBase modelTrain batch sizeBest val mean@16Best checkpointFinal val mean@16Final checkpoint
ChemistrySDPO-TRQwen3-4B3274.91%10074.91%100

Training and validation scores

Raw result files:

  • results/validation_mean16.csv
  • results/training_scores.csv
  • artifacts/config.yaml
  • artifacts/wandb-summary.json

Training Hyperparameters

HyperparameterValue
Base modelQwen/Qwen3-4B
MethodSDPO-TR
Train batch size32
Train max samples3200
Total training steps100
Save frequency10
Test frequency10
Teacher regularizationtrust-region
Teacher update rate0.1
Rollout samples per prompt8
Validation samples per prompt16
Learning rate1e-5
vLLM GPU memory utilization0.8

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "SeongryongJung/Qwen3-4B-Chemistry-SDPO-TR"
tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

Sources

1 source
Hugging FacePrimaryVersion 567d5ca06f64a047943ca8fb626dbe0f13b791dc
huggingface.co/SeongryongJung/Qwen3-4B-Chemistry-SDPO-TR/tree/567d5ca06f64a047943ca8fb626dbe0f13b791dc

Verified Sep 17

Source excerpts

3 excerpts
ReadmeJul 11huggingface.co/SeongryongJung/Qwen3-4B-Chemistry-SDPO-TR

--- license: apache-2.0 base_model: Qwen/Qwen3-4B library_name: transformers tags: - qwen3 - reinforcement-learning - trust-region - sdpo - srpo --- # Qwen3-4B-Chemistry-SDPO-TR This repository contains the Qwen3-4B `Chemistry` `SDPO-TR` run trained with batch size 32. The repository root contains the **best validation checkpoint** (`global_step_100`), selected by validation `mean@16`. The best validation checkpoint is also the final checkpoint, so the repository root is both best and last. ## Performance | Dataset | Method | Base model | Train batch size | Best val mean@16 | Best checkpoint | Final val mean@16 | Final checkpoint | |---|---|---|---:|---:|---:|---:|---:| | Chemistry | SDPO-TR | Qwen3-4B | 32 | 74.91% | 100 | 74.91% | 100 | ![Training and validation scores](results/training_score.png) Raw result files: - `results/validation_mean16.csv` - `results/training_scores.csv` - `artifacts/config.yaml` - `artifacts/wandb-summary.json` ## Training Hyperparameters | Hyperparameter | Value | |---|---:| | Base model | `Qwen/Qwen3-4B` | | Method | `SDPO-TR` | | Train batch size | `32` | | Train max samples | `3200` | | Total training steps | `100` | | Save frequency | `10` | | Test frequency | `10` | | Teacher regularization | `trust-region` | | Teacher update rate | `0.1` | | Rollout samples per prompt | `8` | | Validation samples per prompt | `16` | | Learning rate | `1e-5` | | vLLM GPU memory utilization | `0.8` | ## Usage ```python from transformers import AutoModelForCausalLM, AutoTokenizer repo_id = "SeongryongJung/Qwen3-4B-Chemistry-SDPO-TR" tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True) model = AutoModelForCausalLM.from_pretrained( repo_id, torch_dtype="auto", device_map="auto", trust_remote_code=True, ) ```

NoteSep 17huggingface.co/SeongryongJung/Qwen3-4B-Chemistry-SDPO-TR

Source context: 7 downloads · 0 likes · Pipeline reinforcement-learning · Library transformers · Repo SeongryongJung/Qwen3-4B-Chemistry-SDPO-TR

NoteJul 18huggingface.co/SeongryongJung/Qwen3-4B-Chemistry-SDPO-TR

Source context: 59 downloads · 0 likes · Pipeline reinforcement-learning · Library transformers · Repo SeongryongJung/Qwen3-4B-Chemistry-SDPO-TR