deslop-prompts / README.md
sandeep85's picture
Slop-word prompts with measured Qwen3.5 rates
8b78758 verified
|
Raw History Blame Contribute Delete
3.39 kB
---
license: apache-2.0
language:
- en
pretty_name: Slop-word prompts for steering-vector search
size_categories:
- n<1K
tags:
- steering-vector
- activation-steering
- pivotal-token-search
- pts
- slop
configs:
- config_name: all
data_files: slop_prompts.jsonl
default: true
- config_name: qwen3.5_0.8b_nothink_borderline
data_files: qwen3.5_0.8b_nothink_borderline.jsonl
---
# Slop-word prompts
Short creative-writing prompts and a list of stock literary words ("hum", "whispered", "testament", "vibrant"...), for finding a de-slopping steering vector with [PTS (Pivotal Token Search)](https://github.com/codelion/pts). A prompt counts as a hit when the model's answer uses any listed word.
These prompts produced [LocalLLaMA/Qwen3.5-0.8B-deslop](https://huggingface.co/LocalLLaMA/Qwen3.5-0.8B-deslop).
## Files
| File | Rows | What it is |
|---|---|---|
| `slop_prompts.jsonl` | 600 | Every prompt, with measured rates on Qwen3.5-0.8B and Qwen3.5-9B |
| `qwen3.5_0.8b_nothink_borderline.jsonl` | 200 | Prompts where Qwen3.5-0.8B used a listed word 20–80% of the time: the eval that run used |
| `words.json` | — | The 24 search words, the 11 held-out words, and the regex |
The prompts combine 10 forms (two-sentence story, short poem, diary entry, blog intro, speech...) with 60 everyday subjects (a lighthouse keeper, a night train, a tax auditor...).
## Words
- **Search words (24):** used to find and validate the vector. Only words that are a single token for Qwen3.5 are included, so PTS can tell writing the word apart from choosing to.
- **Held-out words (11):** delve, shimmer, realm, intricate, navigate, embark, embarked, danced, dancing, ancient, serene. They were never searched on, and they test whether a vector reduces slop in general or only the listed words.
- **Dropped:** two character names (Elias, Aria) were removed from the search after "Elias" turned out to dominate the 9B's matches.
## Fields
| Field | Meaning |
|---|---|
| `question` | The prompt |
| `pattern` | Regex for the 24 search words, case-insensitive |
| `category` | The prompt's form |
| `qwen3_5_0_8b_p_slop`, `qwen3_5_9b_p_slop` | Share of 8 answers with a search word |
| `qwen3_5_0_8b_p_held_out`, `qwen3_5_9b_p_held_out` | Share of 8 answers with a held-out word |
Answers were sampled with thinking off, temperature 0.7, top-p 0.8, top-k 20, presence penalty 1.5 (mlx-lm's default 20-token window), up to 400 tokens. Eight samples is coarse (about ±0.17 at a rate of 0.5), so prompts near the 0.2 and 0.8 edges can land either side when measured again.
## Use it
```bash
pip install 'pts[mlx] @ git+https://github.com/codelion/pts'
pts discover --model Qwen/Qwen3.5-0.8B \
--eval hf://LocalLLaMA/deslop-prompts/qwen3.5_0.8b_nothink_borderline.jsonl \
--behavior slop-words --output-dir out/ --num-items 200 --traces-per-item 6 \
--chat-template-args '{"enable_thinking": false}' \
--temperature 0.7 --top-p 0.8 --top-k 20 --presence-penalty 1.5 --max-tokens 400
```
For another model, measure `slop_prompts.jsonl` first and keep its borderline prompts.
## Limitations
- A word list is a narrow proxy for slop. It misses clichés outside the list and says nothing about structure, tone or quality.
- Some listed words are ordinary in plain writing ("shadows", "echo"), so a hit is not always a cliché.
- English only, and short prompts only.