---
format: "aidr-story-markdown/v1"
id: "2937d705c4348bf5a5438bd8aa1e1b70ea5f43b660ed55ae854c4778788020ec"
canonical_url: "https://aidr.today/2937d705?lang=en"
title: "I had Gemini train its own replacement for $9"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-09-17T13:17:16.000Z"
category: "Models"
topics: ["anthropic","claude","multi-agent","open-source"]
source_urls: ["https://www.petervijeh.com/projects/reddit-ner"]
summary: "Gemini 3.1 Pro labeled 4,290 Reddit comments for $9, and a fine-tuned GLiNER model now tags brands, models, and materials locally at 0.83 F1 against those labels. Five of ten training runs produced nothing due to a tensor misread as an attention mask, causing the loss to go flat without an error naming the tensor."
---

# I had Gemini train its own replacement for $9

> [Open the canonical story](<https://aidr.today/2937d705?lang=en>)

**Published:** 2026-09-17T13:17:16.000Z
**Category:** Models
**Topics:** anthropic, claude, multi\-agent, open\-source

## Summary

Gemini 3\.1 Pro labeled 4,290 Reddit comments for $9, and a fine\-tuned GLiNER model now tags brands, models, and materials locally at 0\.83 F1 against those labels\. Five of ten training runs produced nothing due to a tensor misread as an attention mask, causing the loss to go flat without an error naming the tensor\.

## Sources

- [Story source](<https://www.petervijeh.com/projects/reddit-ner>)

