You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

gen_judge_align checkpoint exports

Research checkpoint exports for the gen_judge_align project. This repository is public with gated model-file access and automatic access-request approval. Gating controls access; it does not establish model safety or label correctness.

Contents

Six BF16 Hugging Face-format model exports: three training checkpoints for each of the Qwen3-8B and Qwen3-32B rubric-judge SFT runs.

Run Exported steps
rj_sft_v0:qwen3_8b:rj_sft_v0 87, 174, 261
rj_sft_v0:qwen3_32b:rj_sft_v0 87, 174, 261

Paths preserve the project's original hierarchy:

verl/checkpoints/<run>/exports/global_step_<step>/

Each export includes model shards, configuration, tokenizer files, and its export manifest. This is a multi-checkpoint archive, not a single model at the repository root. Download the desired export subdirectory and load that local directory with compatible model-serving or inference software.

Raw distributed training checkpoints, optimizer state, and files outside exports/ are intentionally excluded.

Intended use and limitations

These models were developed to inspect a question and rubric against supplied rubric principles and return criterion-level flaw explanations. They are research artifacts, not independently certified evaluators. Training references include synthetic model-generated labels, which may contain errors or biases. Checkpoint availability does not imply that a later step is better than an earlier step, that any checkpoint is suitable for deployment, or that the models are reliable on unfamiliar domains.

Consult the project source and experiment records for prompts, training settings, and evaluation protocols. The original base-model terms and notices remain applicable; this archive does not grant additional rights to base-model or source-data content.

Companion repositories

  • Project code: g-h-chen/gen_judge_align on GitHub.
  • Other project data/artifacts: g-h-chen/gen_judge_align on Hugging Face.

Reproducible project sync

After obtaining gated access, use the project's sync utilities from its root:

python3 utils/hf_to_local.py \
  --repo-id g-h-chen/gen_judge_align_ckpt \
  --repo-type model \
  --paths verl/checkpoints

The uploader uses the same relative paths and selects checkpoint exports only.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for g-h-chen/gen_judge_align_ckpt

Base model

Qwen/Qwen3-32B
Finetuned
(595)
this model