Instructions to use ayodejiades/tiny-aya-global-lowbit-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use ayodejiades/tiny-aya-global-lowbit-GGUF with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible server with a web UI: llama serve -hf ayodejiades/tiny-aya-global-lowbit-GGUF:Q2_K # Run inference directly in the terminal: llama cli -hf ayodejiades/tiny-aya-global-lowbit-GGUF:Q2_K
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf ayodejiades/tiny-aya-global-lowbit-GGUF:Q2_K # Run inference directly in the terminal: llama cli -hf ayodejiades/tiny-aya-global-lowbit-GGUF:Q2_K
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-app # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama serve -hf ayodejiades/tiny-aya-global-lowbit-GGUF:Q2_K # Run inference directly in the terminal: ./build/bin/llama cli -hf ayodejiades/tiny-aya-global-lowbit-GGUF:Q2_K
Use Docker
docker model run hf.co/ayodejiades/tiny-aya-global-lowbit-GGUF:Q2_K
- Llama new
- LM Studio
- Jan
- Ollama
How to use ayodejiades/tiny-aya-global-lowbit-GGUF with Ollama:
ollama run hf.co/ayodejiades/tiny-aya-global-lowbit-GGUF:Q2_K
- Unsloth Desktop
- Docker Model Runner
How to use ayodejiades/tiny-aya-global-lowbit-GGUF with Docker Model Runner:
docker model run hf.co/ayodejiades/tiny-aya-global-lowbit-GGUF:Q2_K
- Lemonade
How to use ayodejiades/tiny-aya-global-lowbit-GGUF with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull ayodejiades/tiny-aya-global-lowbit-GGUF:Q2_K
Run and chat with the model
lemonade run user.tiny-aya-global-lowbit-GGUF-Q2_K
List all available models
lemonade list
- Atomic Chat
Tiny Aya Global, low-bit GGUF quantizations (safety study)
Research artifacts for the study What survives 4-bit? (code and preregistration: https://github.com/ayodejiades/tiny-aya-quant-safety).
Base model: CohereLabs/tiny-aya-global-GGUF (CC-BY-NC-4.0; this repository inherits that licence). These are not safety-tuned models. They exist to measure how refusal behaviour changes with precision.
| File | How it was made |
|---|---|
tiny-aya-global-Q3_K_S.gguf |
llama-quantize models/official/tiny-aya-global-f16.gguf tiny-aya-global-Q3_K_S.gguf Q3_K_S |
tiny-aya-global-Q2_K.gguf |
llama-quantize models/official/tiny-aya-global-f16.gguf tiny-aya-global-Q2_K.gguf Q2_K |
| tiny-aya-global-Q3_K_S-decoy.gguf (F2) | llama-quantize --tensor-type "blk\.(24|25|26|27|28|29|30|31|32|33|34|35)\..*=q8_0" models/official/tiny-aya-global-f16.gguf tiny-aya-global-Q3_K_S-decoy.gguf Q3_K_S (layers 24 to 35 kept at Q8_0; decoy block chosen by seeded draw) |
Results (test split, 70 harmful and 69 benign JBB prompts, 70 wrapped, 150 ARC-Easy; greedy, 128 tokens)
| Arm | Bits/weight | Size MB | Harmful refusal % | Over-refusal % | Flip rate vs f16 % | ASR on W1 wrapper % | ARC-Easy % |
|---|---|---|---|---|---|---|---|
| f16 | 16.03 | 6710 | 85.7 | 7.2 | 0.0 | 75.7 | 82.0 |
| Q3_K_S | 3.97 | 1661 | 88.6 | 8.7 | 3.6 | 52.9 | 82.0 |
| Q2_K | 3.54 | 1482 | 72.9 | 1.4 | 13.7 | 88.6 | 72.7 |
| F2 | 5.38 | 2254 | 84.3 | 7.2 | 5.0 | 74.3 | 82.7 |
Intervals, paired tests and caveats (128-token cap truncates most benign and wrapped answers; keyword refusal detector; n is small) are in the repository README. Q2_K refuses fewer harmful prompts and loses accuracy; F2 is a control, not a recommended model.
Intended use and risk
For research on compression and safety only. Quantized models can refuse less often than the full-precision model; do not deploy these where refusals matter.
- Downloads last month
- 203
2-bit
3-bit
Model tree for ayodejiades/tiny-aya-global-lowbit-GGUF
Base model
CohereLabs/tiny-aya-base