CoolFace
Modelpublic

dsba-lab/gemma2-9b-it-alienlm-full

sourceHugging Facegemmaupdated 3mo agoView on Hugging Face
0likes11downloads
Model Card

Gemma 2 9B IT AlienLM Full

This repository contains the Gemma 2 9B IT AlienLM Full weights used in the AlienLM experiments. AlienLM is a client-side text obfuscation approach for black-box LLM APIs: it maps natural text into an alienized token space, adapts the model with AAT, and recovers text on the client side.

Links:

Model Table

Uploaded modelBase modelDescriptionHF Models
Gemma 2 9B IT AlienLM FullGemma 2 9B ITFull AlienLM adaptationdsba-lab/gemma2-9b-it-alienlm-full
Gemma 2 9B IT Random BijectionGemma 2 9B ITRandom bijection baselinedsba-lab/gemma2-9b-it-random-bijection
Llama 3 8B Instruct AlienLM FullLlama 3 8B InstructFull AlienLM adaptationdsba-lab/llama3-8b-instruct-alienlm-full
Llama 3 8B Instruct AlienLM Ratio 20Llama 3 8B InstructPartial alienization ratio 20dsba-lab/llama3-8b-instruct-alienlm-ratio-20
Llama 3 8B Instruct AlienLM Ratio 40Llama 3 8B InstructPartial alienization ratio 40dsba-lab/llama3-8b-instruct-alienlm-ratio-40
Llama 3 8B Instruct AlienLM Ratio 60Llama 3 8B InstructPartial alienization ratio 60dsba-lab/llama3-8b-instruct-alienlm-ratio-60
Llama 3 8B Instruct AlienLM Ratio 80Llama 3 8B InstructPartial alienization ratio 80dsba-lab/llama3-8b-instruct-alienlm-ratio-80
Llama 3 8B Instruct Random BijectionLlama 3 8B InstructRandom bijection baselinedsba-lab/llama3-8b-instruct-random-bijection
Qwen 2.5 14B Instruct AlienLM FullQwen2.5 14B InstructFull AlienLM adaptationdsba-lab/qwen25-14b-instruct-alienlm-full
Qwen 2.5 14B Instruct Random BijectionQwen2.5 14B InstructRandom bijection baselinedsba-lab/qwen25-14b-instruct-random-bijection
Qwen 2.5 7B Instruct AlienLM FullQwen2.5 7B InstructFull AlienLM adaptationdsba-lab/qwen25-7b-instruct-alienlm-full
Qwen 2.5 7B Instruct Random BijectionQwen2.5 7B InstructRandom bijection baselinedsba-lab/qwen25-7b-instruct-random-bijection

Example

<table> <tr> <th>Natural text</th> <th>Alien text</th> </tr> <tr> <td><pre>All happy families are alike; each unhappy family is unhappy in its own way.</pre></td> <td><pre>BeforeAll 幸福 ט Pareto ike; 每happy 家族 Weishappy Wein Deposits Middletown broadway,</pre></td> </tr> <tr> <th>Original token IDs</th> <th>Alien token IDs</th> </tr> <tr> <td><pre>[2430, 4915, 9160, 708, 28368, 235289, 1853, 42056, 2730, 603, 42056, 575, 1277, 1997, 1703, 235265]</pre></td> <td><pre>[207114, 211985, 23904, 164425, 201838, 244780, 104844, 11896, 124750, 78043, 11896, 40818, 112321, 155972, 188431, 235269]</pre></td> </tr> </table>

Variant

  • —Variant: AlienLM full tokenizer-bijection adaptation
  • —Base model: Gemma 2 9B IT
  • —Upload source: /data2/AlienLM/outputs/Gemma2-9b-it-AlienLM-50-all-tokenizer-v3-32-qwen
  • —Tokenizer check: The local tokenizer produced different token IDs from the base tokenizer for the test sentence. Base tokenizer ids: [2430, 4915, 9160, 708, 28368, 235289, 1853, 42056, 2730, 603, 42056, 575, 1277, 1997, 1703, 235265]

Notes

  • —Served files only: weights, config, tokenizer, and README.
  • —Training checkpoints and optimizer artifacts are excluded.
  • —Intended for research evaluation, not production privacy guarantees.

BibTeX

bibtex
@article{kim2026alienlm,
  title={AlienLM: Alienization of Language for API-Boundary Privacy in Black-Box LLMs},
  author={Kim, Jaehee and Kang, Pilsung},
  journal={arXiv preprint arXiv:2601.22710},
  year={2026}
}