Text Generation
MLX
PyTorch
pts
steering-vector
activation-steering
representation-engineering
qwen3_5
Instructions to use LocalLLaMA/Qwen3.5-0.8B-anti-loop with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use LocalLLaMA/Qwen3.5-0.8B-anti-loop with MLX:
# Make sure mlx-lm is installed # pip install --upgrade mlx-lm # if on a CUDA device, also pip install mlx[cuda] # Generate text with mlx-lm from mlx_lm import load, generate model, tokenizer = load("LocalLLaMA/Qwen3.5-0.8B-anti-loop") prompt = "Once upon a time in" text = generate(model, tokenizer, prompt=prompt, verbose=True) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- MLX LM
How to use LocalLLaMA/Qwen3.5-0.8B-anti-loop with MLX LM:
Generate or start a chat session
# Install MLX LM uv tool install mlx-lm # Generate some text mlx_lm.generate --model "LocalLLaMA/Qwen3.5-0.8B-anti-loop" --prompt "Once upon a time"
- Atomic Chat
Download decrease/steering_config.json from LocalLLaMA/Qwen3.5-0.8B-anti-loop: direct link, hf CLI and curl.
- Browser
- Download file 1.49 kB
-
https://huggingface.co/LocalLLaMA/Qwen3.5-0.8B-anti-loop/resolve/main/decrease/steering_config.json
- Command line
-
hf download hf://LocalLLaMA/Qwen3.5-0.8B-anti-loop/decrease/steering_config.json
-
curl -L -o steering_config.json https://huggingface.co/LocalLLaMA/Qwen3.5-0.8B-anti-loop/resolve/main/decrease/steering_config.json
1.49 kB
| { | |
| "format": "pts-steering", | |
| "format_version": "1", | |
| "hook": "block_output", | |
| "base_model": "Qwen/Qwen3.5-0.8B", | |
| "compatible_models": [], | |
| "hidden_size": 1024, | |
| "num_layers": 24, | |
| "default_alpha": 1.0, | |
| "positions": "all", | |
| "behavior": "thinking-loops", | |
| "layers": [ | |
| { | |
| "layer": 15, | |
| "op": "ablate", | |
| "tensor": "layers.15.ablate", | |
| "ref_norm": null | |
| }, | |
| { | |
| "layer": 19, | |
| "op": "ablate", | |
| "tensor": "layers.19.ablate", | |
| "ref_norm": null | |
| } | |
| ], | |
| "provenance": { | |
| "method": "direction removal (ablate), converted to a runtime steering vector", | |
| "source": "Direction recovered from the p50c weight edit in Takehi, Teoh, Zhang and Khattab, 'Sparks of Weight-Space Programming' (2026), made by GPT-6 Astra; logs https://github.com/RikiyaT/lm_understand-logs (CC BY 4.0).", | |
| "recovery": "Top left singular vector of the edit to model.language_model.layers.{15,19}.mlp.down_proj.weight" | |
| }, | |
| "validation": { | |
| "summary": "Reduces thinking loops. Same 100 problems: MATH-500 18 \u2192 46 correct, IFEval 45 \u2192 58. Thinking on, presence penalty 1.5.", | |
| "chat_template_args": { | |
| "enable_thinking": true | |
| }, | |
| "sampling": { | |
| "temperature": 0.6, | |
| "top_p": 0.95, | |
| "top_k": 20, | |
| "presence_penalty": 1.5, | |
| "presence_context": "whole context" | |
| }, | |
| "max_tokens": 32768, | |
| "example_prompt": "What is the capital of Australia, and why isn't it Sydney? Answer in two sentences." | |
| } | |
| } |