Labeling 50k entries (some adult) for $0: Google's filter blocks the model, another host doesn't
Mad-Dog-UA · reddit · 2026-11-04
A developer shares a zero-cost pipeline for labeling 50k web catalogue entries (12 JSON fields each, 3.8k input tokens per call, some adult content): Google AI Studio's free tier returned PROHIBITEDCONTENT on 20/20 entries after adding ordinary moderation sections — a non-adjustable block — while the same Gemma 4 31B via DeepInfra passed 30/30 with identical prompts. Quality versus a Claude Opus reference: Nemotron 3.5 Super VL 73–77% (free), DeepSeek V4.1 Flash 72% ($0.002/item), Sonnet 70%, Gemma 57%; Nemotron's 5k reasoning tokens truncate at maxtokens 8192. Alternating NVIDIA and OpenRouter lanes with 429 backoff yields 250/hour averaged over a day (6k labels/day) for $0 plus a one-time $10 OpenRouter credit. Lesson: measure throughput over 24 hours, not your best two.
More from coding & agent
- Dev open-sources AgentSnow, a minimal coding agent built to demystify Claude Code — damnGruz · 2026-11-04
- Pydantic AI adds Claude Code-style coding tools in one line, countering migration claims — samuelcolvin · 2026-11-04
- Horizon: open-source AI news radar with 12k stars generates daily bilingual briefings — Thysrael · 2026-11-04
- Dev says t3code is "so good" after finally trying it: PRs open in the sidebar — iamsahaj_xyz · 2026-11-04
- Living Brain decouples memory from models: 92.2% LoCoMo, 22.6x fewer tokens — heyshrutimishra · 2026-11-04
- LLMs inherit humanity's bureaucratic instinct: why your codebase fills with nonsense rules — StewartalsopIII · 2026-11-04