Variant

• Variant: Random bijection baseline, seed 42
• Base model: Qwen/Qwen2.5-7B-Instruct
• Local source path used for upload: /data2/AlienLM/outputs/Qwen25-7b-Instruct-random-42
• Weight source used for upload: /data2/AlienLM/outputs/Qwen25-7b-Instruct-random-42
• Tokenizer check: The local tokenizer produced different token IDs from the base tokenizer for the test sentence. Base tokenizer token IDs for the test sentence: [2403, 6247, 8521, 525, 25992, 26, 1817, 42151, 2997, 374, 42151, 304, 1181, 1828, 1616, 13].

Important Limitations

• AlienLM does not provide cryptographic security or formal privacy guarantees.
• The method is deterministic and should be evaluated under the relevant leakage and observer assumptions.
• Safety behavior can differ from the original instruction-tuned model; use this model for research evaluation only.
• Downstream quality depends on task, domain, alienization ratio, and adaptation data.

Tokenization Example

Test sentence:

text
Copy code
All happy families are alike; each unhappy family is unhappy in its own way.

For this repository, the local tokenizer produces these visible token pieces:

text
Copy code
[All, Ġhappy, Ġfamilies, Ġare, Ġalike, ;, Ġeach, Ġunhappy, Ġfamily, Ġis, Ġunhappy, Ġin, Ġits, Ġown, Ġway, .]

The table below records how the same sentence maps to token IDs across the uploaded tokenizers. The visible token pieces may look familiar because AlienLM changes the vocabulary-to-ID mapping; the ID sequence is the important model-facing representation.

Uploaded Files

Only serving-time artifacts were staged for upload:

• added_tokens.json from /data2/AlienLM/outputs/Qwen25-7b-Instruct-random-42/added_tokens.json
• config.json from /data2/AlienLM/outputs/Qwen25-7b-Instruct-random-42/config.json
• generation_config.json from /data2/AlienLM/outputs/Qwen25-7b-Instruct-random-42/generation_config.json
• merges.txt from /data2/AlienLM/outputs/Qwen25-7b-Instruct-random-42/merges.txt
• model-00001-of-00004.safetensors from /data2/AlienLM/outputs/Qwen25-7b-Instruct-random-42/model-00001-of-00004.safetensors
• model-00002-of-00004.safetensors from /data2/AlienLM/outputs/Qwen25-7b-Instruct-random-42/model-00002-of-00004.safetensors
• model-00003-of-00004.safetensors from /data2/AlienLM/outputs/Qwen25-7b-Instruct-random-42/model-00003-of-00004.safetensors
• model-00004-of-00004.safetensors from /data2/AlienLM/outputs/Qwen25-7b-Instruct-random-42/model-00004-of-00004.safetensors
• model.safetensors.index.json from /data2/AlienLM/outputs/Qwen25-7b-Instruct-random-42/model.safetensors.index.json
• special_tokens_map.json from /data2/AlienLM/outputs/Qwen25-7b-Instruct-random-42/special_tokens_map.json
• tokenizer.json from /data2/AlienLM/outputs/Qwen25-7b-Instruct-random-42/tokenizer.json
• tokenizer_config.json from /data2/AlienLM/outputs/Qwen25-7b-Instruct-random-42/tokenizer_config.json
• vocab.json from /data2/AlienLM/outputs/Qwen25-7b-Instruct-random-42/vocab.json

Training-only artifacts such as checkpoint-* directories, trainer_state.json, optimizer states, scheduler states, RNG states, logs, caches, and W&B files were intentionally excluded.

Training Data

The model was adapted on the Magpie instruction and reasoning mixture used in the AlienLM experiments:

• Magpie-Align/Magpie-Pro-300K-Filtered
• Magpie-Align/Magpie-Reasoning-V1-150K

Citation

If you use these weights, please cite the AlienLM paper.