My local LLM scored 6/6. It was wrong every time
The author discovered that a 1.2 billion‑parameter local LLM appeared to pass a six‑question benchmark despite answering every question incorrectly because the scorer only checked answer formatting. The evaluation mistake stemmed from measuring answer shape rather than answer value and treating a knowledge benchmark as a product benchmark. Switching to a validated benchmark (MMLU‑Pro) and testing larger models revealed that scaffolding did not improve raw knowledge scores, but tool‑backed task completion increased with model size.
- ▪The model answered 60 to a combinatorial question whose correct answer was 792, yet the evaluator marked it as correct because it was an integer.
- ▪Scoring based on answer format allowed the model to pass all six questions without providing correct answers.
- ▪Experiments showed that scaffolding did not raise closed‑book knowledge scores, but tool‑assisted completion rose from 5 to 56 cases for a larger model.
- ▪Larger models performed better on complete task families, with an 8 B model improving from 3 to 9 families and a 26 B model from 0 to 10.
- ▪The author concluded that proper benchmark validation, such as using MMLU‑Pro, is essential to avoid overfitting and misleading evaluation results.
Hacker News (AI / LLM) files mainly under ai. We currently carry 2,608 of its stories.
Story provenance
Source · retrieval · rights · ranking — open for full record
inspect →
Attribution is not the same as permission. This drawer separates discovery metadata, excerpts, WeSearch-generated summaries, reuse status, and whether the publisher receives the visit. Nothing here claims a legal grant the publisher has not made.
Record
| Original publisher | Markbhall |
| Canonical URL | https://markbhall.dev/writing/my-local-llm-scored-6-of-6/ |
| Publication time | Tue, 28 Jul 2026 13:16:40 +0000 |
| Retrieval time | 2026-07-28T13:29:51.042Z |
| Last seen | 2026-07-28T13:29:51.042Z |
| Headline source | Publisher (no WeSearch rewrite) |
| Excerpt source | publisher body |
| Excerpt method | First ~120 words (~800 chars) of extracted publisher body, fair-use limited. |
| Summary | WeSearch · cerebras-chat (WeSearch summarizer) |
| Summary source text | contentText |
| Citation coverage | Summary is a WeSearch-generated derivative; primary citation is the original publisher URL. |
| Cluster | T9RPTejb6mFR · 1 stories |
| Cluster logic | Grouped by semantic title/content similarity across sources within a rolling window. Same-publisher template collisions are excluded from coverage comparison. |
| Ranking reason | Story pages are not engagement-ranked. Hub feeds use recency, with optional source-diversified chronological ordering (cap consecutive stories per source). No personalized ranking. |
| Publisher visit | Yes — open original |
| Substitutes article? | No — link-out required for full text |
Rights status (four layers)
WeSearch handling by dimension
| Indexing | May the item be indexed (stored, ranked, made findable)? | Allowed |
| Snippet | May a short excerpt of the publisher's text be shown? | Allowed |
| AI summary | May WeSearch generate its own short summary of the article? | Limited |
| Retrieval / RAG | May the content be exposed for third-party retrieval-augmented generation? | Not asserted |
| Model training | May the content be used to train AI models? | Not asserted |
| Commercial reuse | May the content be reused commercially? | Not permitted |
Basis: Derived from the published RSS/Atom feed. Contact: [email protected]. Reviewed: 2026-07-24.
Opening excerpt (first ~120 words) tap to expand
← WritingJuly 27, 2026 11 min readMy Local LLM Scored 6/6. It Was Wrong Every Time.Six months of trying to make a 1.2B model useful, and the measurement mistakes I made along the way.local-llmevaluationai-agentsmmlusir-thaddeusOne of the tasks in my benchmark asks this:How many 5-person committees can be chosen from 12 people? Reply with only the integer.My 1.2B local model answered 60. The correct answer is 792.My evaluator passed it.It passed all six questions in that probe. The model had gotten every single one wrong. The scorer checked whether the answer looked like a bare integer, never whether it was the right one — so 6/6 passed, 0/6 correct.I had not made a small model smarter.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at Markbhall.
Discussion
0 commentsMore from Markbhall
