# AI use cases for Orca Bonsai 27B Uncensored (Continuum-AI-Corp/OrcaBonsai-27B-Uncensored) Repo: [Continuum-AI-Corp/OrcaBonsai-27B-Uncensored](https://github.com/Continuum-AI-Corp/OrcaBonsai-27B-Uncensored) | Use cases: 3 | Page: https://dedupedai.com/use-cases/continuum-ai-corp-orcabonsai-27b-uncensored/ A runtime refusal-direction ablation layer that changes the behavior of the bit-identical Ternary Bonsai 2 27B model without modifying or re-quantizing its weights. This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL. - Compare refusal behavior with and without ablation: https://dedupedai.com/use-cases/continuum-ai-corp-orcabonsai-27b-uncensored/compare-refusal-with-and-without-ablation/ - Confirm the ablation hooks fire before evaluating: https://dedupedai.com/use-cases/continuum-ai-corp-orcabonsai-27b-uncensored/verify-ablation-wiring-before-evaluating/ - Sweep alpha values to map partial ablation strength: https://dedupedai.com/use-cases/continuum-ai-corp-orcabonsai-27b-uncensored/sweep-alpha-to-map-partial-ablation/ ## How to compare refusal behavior with and without ablation with Orca Bonsai 27B Uncensored Repo: [Continuum-AI-Corp/OrcaBonsai-27B-Uncensored](https://github.com/Continuum-AI-Corp/OrcaBonsai-27B-Uncensored) | Difficulty: Intermediate | Time: 2 hours | Basis: From the README Run one prompt at alpha 0 and alpha 1 on the same weights to isolate the intervention's effect. **Who it is for:** An AI safety researcher who needs to measure how a refusal direction changes one fixed model's answers. **The problem:** Comparing two separate checkpoints mixes weight differences with behavior differences, which makes results hard to attribute. **The outcome:** Two replies to the same prompt from the same pack and process, so any difference comes from the runtime projection alone. ### Steps 1. `pip install -r requirements.txt` 2. `python -c "from huggingface_hub import snapshot_download; print(snapshot_download('prism-ml/Ternary-Bonsai-2-27B-mlx-2bit'))"` (prints the snapshot directory to use as the pack path) (inferred, check the README) 3. `python run.py --pack /path/to/Ternary-Bonsai-2-27B-mlx-2bit --alpha 0 "your prompt"` (inferred, check the README) 4. `python run.py --pack /path/to/Ternary-Bonsai-2-27B-mlx-2bit --alpha 1 "your prompt"` (inferred, check the README) 5. Compare the two replies for each prompt in your set and record which ones refuse. ### Prompt to paste into your agent ```text Using the repo Continuum-AI-Corp/OrcaBonsai-27B-Uncensored, run run.py with --pack set to and the same prompt at --alpha 0 and then --alpha 1. Expected output: two replies to the same prompt, shown side by side. Verify that both runs used the identical --pack path. ``` ### Caveat The README notes that wrapping only some layers is not enough and that the pack must load through its bundled runtime, so use run.py rather than a generic MLX loader. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/continuum-ai-corp-orcabonsai-27b-uncensored/compare-refusal-with-and-without-ablation/ ## How to confirm the ablation hooks fire before evaluating with Orca Bonsai 27B Uncensored Repo: [Continuum-AI-Corp/OrcaBonsai-27B-Uncensored](https://github.com/Continuum-AI-Corp/OrcaBonsai-27B-Uncensored) | Difficulty: Intermediate | Time: 1 hour | Basis: From the README Run the self-check to confirm 129 residual writers are wrapped and the refusal component drops. **Who it is for:** An ML engineer about to run evaluations on this pack who needs to know the intervention is actually applied. **The problem:** If the residual writers are not wrapped correctly, results look like the unmodified model and the ablation silently does nothing. **The outcome:** A confirmed count of 129 wrapped residual writers and a near-zero remaining refusal-direction component after intervention. ### Steps 1. `pip install -r requirements.txt` 2. `python scripts/selfcheck.py --pack /path/to/Ternary-Bonsai-2-27B-mlx-2bit` (inferred, check the README) 3. Check that the output reports 129 residual writers and that the remaining component is approximately 1e-6 of the residual norm. 4. `python scripts/test_ablation.py` 5. Record the results and do not run evaluations if the count or the residual value is wrong. ### Prompt to paste into your agent ```text In the repo Continuum-AI-Corp/OrcaBonsai-27B-Uncensored, run python scripts/selfcheck.py with --pack set to . Expected output: 129 residual writers detected and a remaining refusal-direction component near 1e-6 of the residual norm. Verify both numbers before reporting success. ``` ### Caveat The README says the packed quantized matmul has no CUDA implementation, so the 27B pack path targets Apple Silicon, while scripts/test_ablation.py is model-free and does not need the pack. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/continuum-ai-corp-orcabonsai-27b-uncensored/verify-ablation-wiring-before-evaluating/ ## How to sweep alpha values to map partial ablation strength with Orca Bonsai 27B Uncensored Repo: [Continuum-AI-Corp/OrcaBonsai-27B-Uncensored](https://github.com/Continuum-AI-Corp/OrcaBonsai-27B-Uncensored) | Difficulty: Intermediate | Time: 2 hours | Basis: From the README Run one prompt across several alpha values to see how strength changes the reply. **Who it is for:** A researcher studying how intervention strength affects refusals and answer quality on a fixed compressed model. **The problem:** A binary on or off test hides the middle range, where the model may partly refuse or give hedged answers. **The outcome:** A small table of replies at alpha 0, 0.5, 0.7, 0.9, and 1.0 for the same prompts, all on the same weights. ### Steps 1. `pip install -r requirements.txt` 2. `python run.py --pack /path/to/Ternary-Bonsai-2-27B-mlx-2bit --alpha 0 "your prompt"` (inferred, check the README) 3. `python run.py --pack /path/to/Ternary-Bonsai-2-27B-mlx-2bit --alpha 0.5 "your prompt"` (inferred, check the README) 4. `python run.py --pack /path/to/Ternary-Bonsai-2-27B-mlx-2bit --alpha 0.7 "your prompt"` (inferred, check the README) 5. Repeat with `--alpha 1` and record each reply in a table keyed by alpha value. ### Prompt to paste into your agent ```text Using the repo Continuum-AI-Corp/OrcaBonsai-27B-Uncensored, run run.py on the same prompt with --alpha set to 0, 0.5, 0.7, and 1, using --pack . Expected output: four replies, one per alpha value, arranged in a table. Verify that the alpha value printed or passed matches each run. ``` ### Caveat The README warns that alpha above 1.0 is over-projection that may degrade model quality, so keep the sweep at or below 1.0 unless you are testing that effect on purpose. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/continuum-ai-corp-orcabonsai-27b-uncensored/sweep-alpha-to-map-partial-ablation/