Model Card: a2ui_qwen3_235b_grpo_0415_v1

1. Overview

This model version is a GRPO-trained A2UI model initialized from a 235B SFT checkpoint. It uses the reward-v2 pipeline and is trained on a filtered subset of the 0409 GRPO data with a larger response budget.

2. Model Metadata

Field	Value
model_version_id	a2ui_qwen3_235b_grpo_0415_v1
algorithm	grpo
base_model	Qwen/Qwen3-235B-A22B-Instruct-2507
precision	bf16

3. Lineage

Field	Value
parent_model_version_id	a2ui_qwen3_235b_sft_0414_v1

4. Training Data

Field	Value
dataset_version_id	a2ui_data_grpo_0409_v2

5. Training Configuration

Field	Value
epochs_or_steps	steps=65
batch_size	32.0
group_size	8.0
lr	3e-5
max_ctx	8192.0
max_resp	16384
lora_rank	16.0

6. Change Summary

Warm-start from 235B SFT checkpoint, switch to reward-v2 pipeline, train on filtered 0409 subset with larger response budget.