m2v-qwen25-as-an-embedding-test-1024d

🧪 First Model2Vec with Instruction-Awareness Across 7 Languages

...but that's the only feature, and it's limited!

---

⚠️ Important Disclaimer

This model was an experiment that didn't go as planned.

We tried to distill Qwen2.5-1.5B-Instruct (an instruction-tuned LLM) into static embeddings using Model2Vec, hoping to preserve instruction-awareness. While we did achieve some instruction-awareness in controlled tests, the model performs poorly on standard benchmarks.

We're keeping this model public for transparency and to help others avoid the same pitfalls.

---

📉 MTEB Benchmark Results (The Hard Truth)

Comparison with Better Alternatives

---

🤔 What Went Wrong?

The Hypothesis

We thought: "If we distill an instruction-tuned LLM instead of a base model, the static embeddings might preserve instruction-awareness!"

The Reality

1. Instruction-awareness partially works in controlled tests (96-99% on our custom benchmarks)
2. But standard semantic understanding collapsed - the model doesn't discriminate well between similar and different concepts
3. Embedding collapse: Similar and dissimilar pairs get almost the same similarity scores

The Lesson Learned

Model2Vec distillation works best with models designed for embeddings (like BGE-M3, Qwen3-Embedding), not instruction-tuned LLMs. The instruction-following capability doesn't translate well to static embeddings.

---

📊 What This Model CAN Do (Limited)

In our controlled monolingual tests, the model showed instruction-awareness:

# This works (instruction understood)
"Explain neural networks" → "Neural networks tutorial guide"  # High similarity

# But general semantic understanding fails
"The cat sits on the mat" vs "A dog runs in the park"  # Should be different, but scores are similar

Languages tested: EN, FR, ES, DE, ZH, AR, RU (monolingual only, no cross-lingual)

---

🎯 Should You Use This Model?

❌ NO, unless:

• You're doing research on Model2Vec limitations
• You want to understand what happens when distilling instruction-tuned LLMs
• You're specifically testing instruction-awareness (not general embeddings)

✅ Use these instead:

• tss-deposium/m2v-bge-m3-1024d - Our best Model2Vec (0.47 MTEB)
• tss-deposium/m2v-qwen3-embedding-0.6b-1024d - Smaller alternative (0.42 MTEB)
• BAAI/bge-m3 - Full transformer, best quality

---

🔧 Technical Details

• Base Model: Qwen/Qwen2.5-1.5B-Instruct (NOT an embedding model!)
• Distillation: Model2Vec static embeddings
• Dimensions: 1024D
• Size: 65MB
• Problem: Base model wasn't designed for embeddings

---

📚 What We Learned

1. Don't distill instruction-tuned LLMs for embeddings - use embedding-specific models
2. Custom benchmarks can be misleading - always validate with standard benchmarks (MTEB)
3. Instruction-awareness ≠ semantic understanding - they're different capabilities
4. Transparency matters - publishing failures helps the community

---

🔗 Better Alternatives (Our Models)

---

📖 Citation

If you reference this model (as an example of what NOT to do):

@misc{m2v-qwen25-test,
  author = {Nicolas Geysse - TSS Deposium},
  title = {m2v-qwen25-as-an-embedding-test: An Experiment in Distilling Instruction-Tuned LLMs},
  year = {2025},
  note = {Experimental model with poor MTEB performance (21.57%). Published for transparency.},
  url = {https://huggingface.co/tss-deposium/qwen25-deposium-1024d}
}

---

Built by Nicolas Geysse - TSS Deposium

"We tried something new. It didn't work great. But we learned, and now you can too!" 🚀

Our Better Models • Report Issues