Metric	Value
Model	Svngoku/qwen3-black-mirror (https://huggingface.co/Svngoku/qwen3-black-mirror)
Base	Qwen3-0.6B-Base (4-bit)
Data	33 episodes across 7 seasons, auto-merged
Epochs	3
Steps	15
Training time	25 seconds
Loss	3.08 -> 2.09 (epoch 2) -> 2.54 (epoch 3)
Upload	Merged 16-bit model (1.19 GB)
The model was merged (LoRA weights folded into the base model) and pushed as a standalone model to Svngoku/qwen3-black-mirror. The inference test shows it's generating coherent text with some thematic influence (surveillance, privacy, ratings -- very Black Mirror).
Artifacts created:

scripts/continued-pretraining.py -- custom script with multi-split support
Svngoku/black-mirror-training (https://huggingface.co/datasets/Svngoku/black-mirror-training) -- HF dataset repo hosting the script
Svngoku/qwen3-black-mirror (https://huggingface.co/Svngoku/qwen3-black-mirror) -- the trained model
Svngoku/qwen3-black-mirror-test (https://huggingface.co/Svngoku/qwen3-black-mirror-test) -- dry-run LoRA adapter

Uploaded finetuned model

Developed by: Svngoku
License: apache-2.0
Finetuned from model: unsloth/Qwen3-0.6B-Base-unsloth-bnb-4bit

This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.

Svngoku/qwen3-black-mirror