Satoshi Akiyama publishes local-model experiments under the Bigbonus GitHub identity. His context-window investigation examines how runtime configuration and evaluation code can produce misleading conclusions about a model.[1][2]
Start with Ollama served my 40k-context model at 4k, silently →
Links
GitHub Contributions
Akiyama found a Qwen model advertising 40,960 tokens running with a 4,096-token context on his Ollama setup. Comparing the model description, running-server context and prompt_eval_count exposed truncation that could still return HTTP 200. He also corrected a scoring bug that classified empty, reasoning-budget-exhausted responses as confabulations. The report incorporates upstream explanations and limits its verified findings to Ollama 0.32.9 on Windows through /api/chat. It teaches measurement hygiene rather than ranking models or proving that knowledge distillation succeeds.[2]
Tools, writing & talks
Writing, research & talks
Start here
A reproducible report and probe for applied context size and silent truncation.
Read the two one-line checks, then the version and platform limits. Running the probe requires a local Ollama model and command-line familiarity; compare runtime token counts with what you sent.
Author and repository owner[2]Original source