unsloth-jobs-training

name: unsloth-jobs-training description: Submit a LoRA fine-tuning job to Hugging Face Jobs using Unsloth optimizations on LFM2.5-1.2B-Instruct. Use when the user wants to train or fine-tune a model on HF Jobs cloud GPUs. argument-hint: [dataset-id] [output-repo-id] disable-model-invocation: false allowed-tools: Bash

Submit a LoRA fine-tuning job to Hugging Face Jobs using Unsloth.

Arguments

$ARGUMENTS should contain the dataset ID and optionally the output repo ID. If not provided, ask the user for: dataset ID, output repo ID, and run name.

Default Configuration (proven from LFM2.5 mobile-actions training)

• Base model: LiquidAI/LFM2.5-1.2B-Instruct
• LoRA rank/alpha: 16
• Batch size: 2 (effective 8 with grad accumulation 4)
• Learning rate: 2e-4, linear decay
• Epochs: 3
• Max seq length: 2048
• Eval split: 2% from training data
• Hardware: a10g-large (24GB VRAM)
• Timeout: 4h
• Trackio space: kshitijthakkar/trackio

Step 1: Confirm authentication

hf whoami

Step 2: Submit the job

Run this command (substitute <DATASET_ID>, <OUTPUT_REPO>, <RUN_NAME>):

hf jobs uv run \
  --flavor a10g-large \
  --timeout 4h \
  --secrets HF_TOKEN \
  "C:\Users\kshit\.claude\plugins\cache\huggingface-skills\hugging-face-model-trainer\3f4f55d6264b\scripts\unsloth_sft_example.py" \
  -- \
  --dataset <DATASET_ID> \
  --num-epochs 3 \
  --lora-r 16 \
  --lora-alpha 16 \
  --batch-size 2 \
  --gradient-accumulation 4 \
  --learning-rate 2e-4 \
  --max-seq-length 2048 \
  --eval-split 0.02 \
  --trackio-space kshitijthakkar/trackio \
  --run-name <RUN_NAME> \
  --seed 3407 \
  --output-repo <OUTPUT_REPO>

Step 3: Report back to user

After submission, provide:

• Job ID (from command output)
• Job URL: https://huggingface.co/jobs/kshitijthakkar/<job-id>
• Trackio dashboard: https://huggingface.co/spaces/kshitijthakkar/trackio
• Estimated completion: ~2-3 hours for 30k samples, 3 epochs on A10G

Check job status

PYTHONUTF8=1 hf jobs inspect <job-id>

CLI syntax rules (critical)

• All HF Jobs flags (--flavor, --timeout, --secrets) MUST come before the script path
• Use -- (double dash) to separate HF Jobs flags from script arguments
• PYTHONUTF8=1 prefix avoids Windows encoding errors in log output
• Local script files are uploaded automatically by the CLI

Dataset requirements

Dataset must have a messages, conversations, or conversation column with lists of {role, content} dicts. The script handles column renaming and ChatML formatting automatically.

Validated example (working job)

hf jobs uv run \
  --flavor a10g-large \
  --timeout 4h \
  --secrets HF_TOKEN \
  "C:\Users\kshit\.claude\plugins\cache\huggingface-skills\hugging-face-model-trainer\3f4f55d6264b\scripts\unsloth_sft_example.py" \
  -- \
  --dataset kshitijthakkar/liquidchat-lora-dataset \
  --num-epochs 3 \
  --lora-r 16 --lora-alpha 16 \
  --batch-size 2 --gradient-accumulation 4 \
  --learning-rate 2e-4 --max-seq-length 2048 \
  --eval-split 0.02 \
  --trackio-space kshitijthakkar/trackio \
  --run-name liquidchat-lora-v1 \
  --seed 3407 \
  --output-repo kshitijthakkar/LFM2.5-1.2B-Instruct-liquidchat-lora

Job ID: 699d9b0f52d1c53b7df7d678 — GPU: A10G — Started: 2026-02-24

Hardware tiers and batch sizes

Gradient accumulation is always 4. Adjust batch size to match VRAM.

Troubleshooting

Next step after training

Run /lora-to-cactus-hub to merge adapters and convert to Cactus format for mobile deployment.

Supporting files

This skill includes reference material and alternative scripts:

• TRAINING_GUIDE.md — Full training guide: hardware selection, cost estimates, step-by-step workflow, and post-training checklist
• TRAINING_TEST_RESULTS.md — Validated dataset stats, configuration table, and performance expectations
• scripts/train_liquidchat_lora.py — Self-contained training script (alternative to the plugin script, useful for local GPU runs)
• scripts/run_hf_jobs_training.py — HF Jobs-compatible PEP 723 script with inline dependencies
• scripts/submit_training_job.py — Python API submission helper (reference for programmatic job submission)