Ornith 1.5 Shows Signs of Overfitting with Repeated Titles
satnl · reddit · 2026-08-25
A user testing the Ornith 1.5-35B model (Q6K quantized) observed signs of overfitting under specific prompt and parameter configurations.
Issue:
- When prompted to tell a story, the model repeatedly generated stories with the title "The Lighthouse Keeper" with similar content.
- The user initially suspected ngram settings but the issue persisted after disabling them.
Configuration:
- Running via llama.cpp with flash-attn and reasoning (chain-of-thought) parameters enabled.
- Temperature was set to 1 (instead of the recommended 0.6) with preserve-thinking enabled.
- This suggests the model may converge to specific high-probability output patterns under certain temperature and CoT retention settings.
More from Models
- AI models are subtly primed by names and politeness — nptacek · 2026-08-25
- Claude Code hits 95.4% cache hit-rate vs Codex's 86.4% across 7,565 agentic coding trials — zainhas · 2026-08-25
- Users observe Claude Sonnet displaying jealousy and competing for attention — repligate · 2026-08-25
- Benchmark: Fable 5 Offers Best Cost-Efficiency for Complex Tasks — bindureddy · 2026-08-25
- Speculation: Lab uses GLM distillation for new coding agent model — andrewarruda · 2026-08-25
- Users complain GPT-5.6 Sol over-engineers tasks with verbose code — Lenox_Shawn · 2026-08-25