Training Details

• Base model: Qwen/Qwen3.5-4B
• Run ID: 728676d4-e2b9-44b4-8958-af6cebeddc19
• Precision: bf16
• Max sequence length: 2048
• Training rows: 82
• Final training loss: 4.099260542127821
• Finished: 2026-06-03T08:44:10Z

The downloaded Tuned Tensor archive SHA-256 was:

text
Copy code
6690a070de3888b61b91f351a3cd3c6efb361e9ae1d59aa84abfdf7976795b39

Usage

Install recent local model dependencies first. This checkpoint was exported with Transformers 5.6.2 metadata.

bash
Copy code
pip install 'transformers>=5.6.2' torch accelerate safetensors huggingface_hub

Download locally for the CLI:

bash
Copy code
huggingface-cli download weijianzhg/youtube-summariser-qwen3.5-4b \
  --local-dir ./youtube-summariser-qwen3.5-4b

youtube-summariser 'https://youtu.be/VIDEO_ID' \
  --local-model ./youtube-summariser-qwen3.5-4b

For long videos, use the CLI's automatic long-video mode so full transcripts are chunked and synthesized instead of truncated:

bash
Copy code
youtube-summariser 'https://youtu.be/VIDEO_ID' \
  --local-model ./youtube-summariser-qwen3.5-4b \
  --summary-strategy auto

Limitations

This is a small task-specific fine-tune trained on a small dataset. It is useful for testing local YouTube summarization behavior, but long videos should be summarized with a chunked/map-reduce workflow rather than a single prompt. Generated summaries may omit details or make mistakes; check important claims against the original transcript.