You can deploy this model and start earning money today!

ztiganj/
verl-edits-counterfact_baseline_format_correctness-step200

Name: verl-edits-counterfact_baseline_format_correctness-step200
Availability: InStock

safetensors

llama

llama3.1

Views

No views yet

ztiganj/verl-edits-counterfact_baseline_format_correctness-step200

Compact Hugging Face inference export from a VERL FSDP actor checkpoint.

Source

Modal volume: verl-setup-outputs
Source actor directory: /checkpoints/verl_edits/counterfact_baseline_format_correctness/global_step_200/actor
Run: counterfact_baseline_format_correctness
Global step: 200

Conversion

Converted with the official VERL FSDP model merger:

python -m verl.model_merger merge --backend fsdp --local_dir <actor_dir> --target_dir <target_dir>

The original training step stored FSDP actor shards plus extra state. This repository keeps only the merged Hugging Face inference checkpoint, tokenizer/config files, and a migration manifest.

Migration Status

This repository was made public after verifying the compact export and checking that projected public Hugging Face storage remained below the configured guardrail.