Qwen3.5-0.8B-anti-loop / decrease /steering_config.json
sandeep85's picture
Open on a prompt that shows the difference
4ab4634 verified
Raw History Blame Contribute Delete
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."
}
}