A local 8B model invented 7.2% of the numbers in an ML pipeline summary
Former_Appointment84 · reddit · 2026-07-23
A Reddit user reports measuring how often a local 8B model invents numbers while writing over an ML pipeline output, and found a 7.2% fabrication rate.
The setup used a checker that matched every generated number against a ground-truth pool built from an XGBoost → SHAP → optimizer pipeline. On 30 seeds with Llama 3.1 8B Instruct Q4KM via llama.cpp, the model wrote 138 numbers and fabricated 10 of them; a frontier API control wrote 537 numbers with none fabricated.
The surprising failure mode was not small numerical drift but invented structures, such as ROI tables or budget splits that never existed in the pipeline. The author notes the result is only one model, one quantization level, and one prompt chain, and asks whether the issue is quantization damage or simply the limits of an 8B model.
More from Research
- Sentence Transformers 5.6.1 fixes a silent flash-attention regression in RoBERTa models — tomaarsen · 2026-07-23
- Sentence Transformers packed position IDs wrong for RoBERTa models under flash attention — tomaarsen · 2026-07-23
- BAAI/bge-m3 lost accuracy under flash attention until the Sentence Transformers patch — tomaarsen · 2026-07-23
- Monthly derivative-model shares now tilt heavily toward China — natolambert · 2026-07-23
- Daily open-model dashboard shows China still far ahead of the US — natolambert · 2026-07-23
- Teaching notebook builds a neural network from scratch with linear algebra and the chain rule — GeostatsGuy · 2026-07-23