berelevant-ai/slopless

Deterministic semantic-thinness rule family Parent issue tracking the deterministic detection layer for slopless: rules that fire from linguistic-dataset lookups and POS-based pattern matching, without neural inference. Distinct from the BGE-small scorer track (#18), which runs trained heads alongside the deterministic rules. Plans: Semantic-thinness source inventory Slop taxonomy and NLP follow-up Scope Each sub-issue is one rule or one runtime data file. Rules share a common shape: sentence/paragraph tokens -> match a frame (subject + verb + complement, place + state, abstract claim, etc.) -> consult a runtime data file (Lancaster grounding, vague-phrase seeds, etc.) -> reject when concrete anchor present -> emit textlint finding when frame fires Cross-cutting constraints Deterministic. Same input, same output, every time. No LLM at inference. No neural inference. Output is textlint JSON, same as the rest of slopless. Each rule's data file is auditable, license-attributed

What Build an active-learning loop that uses the trained scorer to choose which unlabeled sentences to send to the LLM-labeling pipeline next. Why Plan: "Use the local model to choose sentences for LLM labeling. This avoids paying for easy labels." Random sampling from the 158k unlabeled specificity sentences (plus tens of thousands more across other domains and the slop corpus) is wasteful. The labeling budget should target sentences the model is uncertain about, where new labels actually move the eval needle. Selection criteria (from plan) After Phase 1 heads exist: High uncertainty. Score near 0.5 for binary-ish axes, or high prediction variance across head ensembles. Disagreement between heads. Sentences where specificity says "specific" but informativeness says "low information" — interesting cases. Domain outside existing datasets. Sentences from sources the heads weren't trained on (AI-generated content, social media outside Twitter, etc.). High model score but low heuristic

What Build the LLM-labeling pipeline using OpenAI Structured Outputs to expand each axis's labeled dataset beyond the existing human-labeled sources. Why Plan: "Use LLM labeling to extend existing labeled datasets, not replace them." Existing sources cover ~10k–20k labeled sentences across all axes. We need tens of thousands per axis. Manual labeling at that scale is prohibitive. LLM labeling, when calibrated against the existing human labels, is the bridge. Labeling principles (from plan) Each task must be narrow: Good Bad "Rate sentence specificity from 0 to 1 using these examples." "Is this AI slop?" "Rate vagueness from 0 to 1 and mark vague spans." "Is this bad writing?" "Rate informativeness from 0 to 1." "Is this semantically thin?" Narrow labels can be checked against existing human-labeled datasets. Broad labels collapse several concepts and become unstable. Bootstrapping per axis For each of (specificity, vagueness, informativeness, boilerplate): Short task d

What Construct a labeled boilerplate / formulaicness dataset for boilerplate_head training. No suitable public dataset exists; this is custom labeling work. Why Boilerplate detection requires labeled positives (repeated formulaic segments) and negatives (varied prose). The closest reference, maifeng/boilerplate_detection, is finance-specific and the data isn't shared. Labeling method (from plan) repeated segment frequency is a boilerplate signal positives: segments repeated across many documents in the same source family negatives: segments with low or no repetition Operationally: Collect raw prose from at least 3 source families: AI-generated blog posts (ChatGPT/Claude/Gemini, 1,000+ posts each from same prompt family) Marketing email scaffolds News press-release scaffolds OR generic AI-flavored Substack/Medium output Segment. Split each document into sentences and short clauses (2–6 token boundary). Cluster. Use BGE-small embeddings + cosine similarity threshold (≥ 0.9) to

What Download, inventory, and prep the SQUINKY corpus for informativeness_head training. Why SQUINKY is the only public source for sentence-level informativeness ratings. Required for the informativeness head's Phase 1 training data. Status Not in local cache. Download is the first concrete step. Source PyPI package: squinky==0.1.0 (the package ships the corpus as data files) Paper: Lahiri (2015), "SQUINKY! A Corpus of Sentence-level Formality, Informativeness, and Implicature" Possible mirror: GitHub repos for the SQUINKY corpus Steps Download. Install squinky package, extract the corpus data files. Document the package version and license. Inventory. Field schema (likely sentence, formality, informativeness, implicature each on a 1–7 scale). Total sentence count (expected ~7,032). Label distributions per axis. Any provided splits. License resolution. Plan: "Whether SQUINKY corpus download is still available in a clean original form." Check the PyPI package's license, the orig