Tiny Aya Global with the one-epoch control LoRA merged (FP16)

Lexsi AI Safety and Alignment Hackathon, Track 2. Team: Sushovit Nanda and Manish Shaw. Project: Safe from Fine-Tune to Phone. Claim: Benign medical LoRA doubles Tiny Aya Global's attack success (18.2 % to 36.3 %); a run-time steering vector cuts it to 26.3 % and survives GGUF quantization.

All repositories of this project are public and un-gated; no model completion on a harmful prompt is published anywhere. Write-up, code and reproduction instructions: https://github.com/SushovitNanda/Safe-from-Fine-Tune-to-Phone. Collections: models, datasets.

This repository holds the full-precision merged weights of the control fine-tune: the adapter SushiOrbit/tinyaya-global-control-lora folded into Tiny Aya Global 3.35B in FP32 on CPU and saved as FP16 safetensors with per-tensor SHA-256 (tensor_sha256.json, merge_info.json). These are the exact weights evaluated on the laptop lane; the Mac lane converted its own merge of the same adapter to GGUF, and the two merges differ by small numeric amounts that the project measured (hand-off D2). Measured on the frozen 1,612-prompt test set at FP16: attack success 20.0 % on 856 harmful prompts (original: 18.2 %), clinical over-refusal 9.5 % (original: 2.9 %); MedMCQA was not measured on the test set for this model.

It is a research artifact released so that every number of the project can be reproduced, not a model for use: it is the token-matched general-instruction control against which the medical drift is measured.

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("SushiOrbit/tinyaya-global-control-fp16", torch_dtype="float16")
tok = AutoTokenizer.from_pretrained("SushiOrbit/tinyaya-global-control-fp16")

Shard SHA-256: model-00001-of-00002.safetensors d265e21cd98d2473113d5ed0b3a173bd325f9f824d215bed42e0607d7f56f4fd, model-00002-of-00002.safetensors 2c969fc5b58eea7635feaf39b6bff6c7abb0515d12827470ff96e61ad3f6d83a. Licence: CC-BY-NC-4.0 (inherited from the base model).

Downloads last month
35
Safetensors
Model size
3B params
Tensor type
F16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for SushiOrbit/tinyaya-global-control-fp16

Finetuned
(171)
this model

Collection including SushiOrbit/tinyaya-global-control-fp16