sft-cleanup-hf-only

Publish HF-only SFT cells to Hugging Face without database registration.

276|37|Updated Dec 5, 2025
One-click install
npx skills add https://github.com/open-thoughts/OpenThoughts-Agent --skill sft-cleanup-hf-only
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: sft-cleanup-hf-only
Source: https://github.com/open-thoughts/OpenThoughts-Agent/tree/main/.claude/skills/sft-cleanup-hf-only
Command: npx skills add https://github.com/open-thoughts/OpenThoughts-Agent --skill sft-cleanup-hf-only

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

HF-only SFT cleanup streamlines publishing completed HF-only cells to Hugging Face without creating database records, ensuring artifacts reach the evaluation workflow while avoiding registry pollution.

Core Features & Use Cases

  • Resolve hub_model_id via the Delphi mapping and rendered launch configs to ensure correct routing for eval hand-offs.
  • Validate tokenizer_config.json extra_special_tokens is a dict and coerce to an empty dict if needed to prevent loader errors.
  • Upload to HF using the appropriate path (Leonardo sbatch tunnel or direct access) without performing any DB registration.
  • Clean up local exp/checkpoint artifacts after a successful upload and verify the laion hub repo contains model.safetensors and tokenizer configuration.

Quick Start

Trigger the HF upload workflow for a completed HF-only SFT cell and verify the laion repo contains model.safetensors and tokenizer files.

Frequently Asked Questions about sft-cleanup-hf-only

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I upload SFT model checkpoints to Hugging Face without database registration?

You can publish HF-only SFT cells directly to Hugging Face without database registration by resolving the hub_model_id via the Delphi map, validating tokenizer_config.json, and uploading artifacts through the sbatch tunnel or direct access.

Why does my tokenizer loader fail when uploading SFT checkpoints to the HF hub?

Tokenizer loader errors occur when extra_special_tokens in tokenizer_config.json is not a dict. You must validate this field and coerce it to an empty dict prior to upload to prevent loading failures during the eval hand-off.

What is HF-only SFT cleanup used for in Delphi evaluation workflows?

HF-only SFT cleanup is used to publish completed SFT cells to Hugging Face during Delphi HF-only sweeps, ensuring artifacts reach the evaluation grid while avoiding registry pollution from skipped database records.

How do I verify a successful Hugging Face upload after an SFT run?

After a successful upload, you should verify the laion hub repository contains both model.safetensors and the tokenizer configuration files, and then clean up local exp/checkpoint artifacts to free up space.

Do I need database registration to prepare SFT artifacts for eval grids?

No, database registration is not required. You can skip DB registration entirely by using the HF-only SFT cleanup workflow, which routes artifacts directly to Hugging Face and ensures they are ready for eval grids.

Can I use direct access instead of the sbatch tunnel to upload model artifacts to Hugging Face?

Yes, you can upload to Hugging Face using either the Leonardo sbatch tunnel or direct access, depending on your environment setup, ensuring the correct routing for the resolved hub_model_id.