Uploaded October 2026 | Updated November 2026, 1 day ago
Fine-tune Embedding Gemma 2 (EmbeddingGemma 2) on your own data with Unsloth and LoRA in a free Colab notebook. I show how fine-tuning an embedding model differs from fine-tuning an LLM, then train Google's multimodal embedding model on two tasks: environmental sounds (ESC-50 top-1 24.5% to 65.8%) and search over my own YouTube transcripts (66.3% to 75.0% top-1 on videos it never saw), and check what it costs the rest of the model.
You'll learn how to build question-passage training pairs, why in-batch negatives (MultipleNegativesRankingLoss) need a no-duplicates sampler, why training and search prompts must match, how LoRA works on a shared multimodal backbone, how to evaluate on unseen data, and a reload bug that silently breaks audio adapters.
If you have fine-tuned embedding models for your own applications, share what worked for you in the comments.
Links:
Colab notebook: colab.research.google.com/drive/1xq-84DLZsqteEzL52Qb__47Y_yGckeyB?usp=sharing
The first EmbeddingGemma 2 video: [EMBEDDINGGEMMA 2 VIDEO LINK]
Unsloth EmbeddingGemma 2 guide: unsloth.ai/docs/models/embeddinggemma-2
Unsloth notebooks: github.com/unslothai/notebooks
Model card: huggingface.co/google/embeddinggemma-2
ESC-50 dataset: github.com/karolpiczak/ESC-50
Sentence Transformers losses: sbert.net/docs/package_reference/sentence_transformer/losses.html
00:00 - Fine-Tuning EmbeddingGemma 2 on Your Own Data
01:03 - Embedding Fine-Tuning vs LLM Fine-Tuning
02:30 - In-Batch Negatives (MultipleNegativesRankingLoss)
03:37 - The Duplicate Trap: No-Duplicates Sampler
04:45 - LoRA on a Shared Multimodal Backbone
06:30 - Notebook: ESC-50 Sounds with Unsloth on an A100
09:32 - Side Effects on Photo, Voice and Text Search
10:06 - Gotcha: Reloading an Audio Adapter
10:50 - Fine-Tuning on YouTube Transcripts
12:04 - Three Things to Get Right
Fine-tune Embedding Gemma 2 (EmbeddingGemma 2) on your own data with Unsloth and LoRA in a free Colab notebook. I show how fine-tuning an embedding model differs from fine-tuning an LLM, then train Google's multimodal embedding model on two tasks: environmental sounds (ESC-50 top-1 24.5% to 65.8%) and search over my own YouTube transcripts (66.3% to 75.0% top-1 on videos it never saw), and check what it costs the rest of the model.
You'll learn how to build question-passage training pairs, why in-batch negatives (MultipleNegativesRankingLoss) need a no-duplicates sampler, why training and search prompts must match, how LoRA works on a shared multimodal backbone, how to evaluate on unseen data, and a reload bug that silently breaks audio adapters.
If you have fine-tuned embedding models for your own applications, share what worked for you in the comments.
Links:
Colab notebook: colab.research.google.com/drive/1xq-84DLZsqteEzL52Qb__47Y_yGckeyB?usp=sharing
The first EmbeddingGemma 2 video: [EMBEDDINGGEMMA 2 VIDEO LINK]
Unsloth EmbeddingGemma 2 guide: unsloth.ai/docs/models/embeddinggemma-2
Unsloth notebooks: github.com/unslothai/notebooks
Model card: huggingface.co/google/embeddinggemma-2
ESC-50 dataset: github.com/karolpiczak/ESC-50
Sentence Transformers losses: sbert.net/docs/package_reference/sentence_transformer/losses.html
00:00 - Fine-Tuning EmbeddingGemma 2 on Your Own Data
01:03 - Embedding Fine-Tuning vs LLM Fine-Tuning
02:30 - In-Batch Negatives (MultipleNegativesRankingLoss)
03:37 - The Duplicate Trap: No-Duplicates Sampler
04:45 - LoRA on a Shared Multimodal Backbone
06:30 - Notebook: ESC-50 Sounds with Unsloth on an A100
09:32 - Side Effects on Photo, Voice and Text Search
10:06 - Gotcha: Reloading an Audio Adapter
10:50 - Fine-Tuning on YouTube Transcripts
12:04 - Three Things to Get Right






![98% of Qwen - Bonsai 2
Bonsai 2 from PrismML squeezes Qwen3.8 27B into a 6 GB file by retraining every weight to −1, 0 or +1, and PrismML claims it keeps 98% of the full models performance. On short tasks it matched the full model in my tests. Then I gave it a real job in an agentic loop: build an ISS tracker on a 3D globe. The full Qwen model built a working tracker. Bonsai never wrote a line of the app. It ran the exact same search 114 times and got the same answer every time.
Compression holds up for chat and quick coding help on a laptop. For long-running agentic tasks, test it on your own tasks before you trust the benchmarks.
Full comparison: [link to the full video]
#Bonsai2 #Qwen #LocalLLM #AIAgents #OpenSourceAI 98% of Qwen - Bonsai 2](https://i.ytimg.com/vi/TOmMp4SU144/mqdefault.jpg)



