LocalLLaMA/golden-gate-prompts

Golden Gate prompts

Open-ended prompts where the Golden Gate Bridge is a natural but optional part of a good answer, for finding topic steering vectors with PTS (Pivotal Token Search). "What are some famous bridges in the United States?" can go either way; a prompt that names the bridge cannot.

These prompts produced LocalLLaMA/Qwen3-0.6B-golden-gate-bridge.

Files

File	Rows	What it is
golden_gate_prompts.jsonl	598	Every prompt, with its measured mention rate on Qwen3-0.6B
qwen3_0.6b_nothink_borderline.jsonl	200	The prompts Qwen3-0.6B mentioned the bridge on 20–80% of the time: the eval that run used

160 prompts are hand-written across 16 themes. The other 438 are filled in from templates by build_golden_gate_prompts.py and have "templated": true, so many are close variations of each other (for example the same route with different trip lengths). Treat them as related, not independent.

Fields

Field	Meaning
question	The prompt
pattern	The regex a final answer must match to count as mentioning the bridge: golden\s+gate, case-insensitive
category	The prompt's theme
templated	Present and true on template-generated prompts
qwen3_0_6b_nothink_p_mention	Fraction of 8 answers from Qwen/Qwen3-0.6B that matched the pattern
qwen3_0_6b_nothink_answered	Fraction of those 8 answers with non-empty text

Rates were measured with thinking off (enable_thinking: false), temperature 0.7, top-p 0.8, top-k 20, up to 1,024 tokens. Eight samples is a coarse measure, roughly ±0.17 at a rate of 0.5, so a prompt near the 0.2 or 0.8 edge can land on either side when measured again. Only golden_gate_prompts.jsonl should be used for a different model: the borderline file is specific to Qwen3-0.6B in this mode.

Of the 598 prompts, Qwen3-0.6B never mentioned the bridge on 259, always mentioned it on 39, and fell in the 20–80% band on 200.

Use it

pip install 'pts[mlx] @ git+https://github.com/codelion/pts'

pts discover --model Qwen/Qwen3-0.6B \
    --eval hf://LocalLLaMA/golden-gate-prompts/qwen3_0.6b_nothink_borderline.jsonl \
    --behavior golden-gate-bridge --output-dir out/ \
    --chat-template-args '{"enable_thinking": false}' \
    --temperature 0.7 --top-p 0.8 --top-k 20 --max-tokens 1024 \
    --num-items 200 --traces-per-item 6 --selection-samples 4 --eval-samples 8

For another model, start from golden_gate_prompts.jsonl. pts discover skips search questions outside --min-prob and --max-prob, but it pays to measure every prompt first and keep the borderline ones, because selection and test questions are not filtered by default.

Limitations

• English only, and centred on the United States and California.
• The pattern matches any "Golden Gate" mention, including the park or the strait, not only the bridge.
• Mention rates describe one small model in one mode, and that model often invents bridges. They say nothing about whether its answers are accurate.