RULER 128K
RULER 128K - long context retrieval benchmark evaluated at 128K tokens. As evaluated on the LFM3-1.7B card: model served via vLLM in BF16 at its full 262,144-token context, temperature 0.6, top_p 0.95, no output limit other than the context length.
Category
Long Context
Score type
percent
Scale
100%
higher is better
Models tested
1
What is RULER 128K?
RULER 128K is a benchmark for reading and reasoning over very long inputs. The official description says: "RULER 128K - long context retrieval benchmark evaluated at 128K tokens". Below we describe in simple words what benchmarks of this type generally examine.
What does it test?
RULER 128K measures reading and reasoning over very long inputs. In detail, it checks:
- finding relevant facts buried in documents hundreds of thousands of tokens long
- retrieving multiple needles hidden across the full context
- reasoning over whole books, codebases or conversation histories
- staying accurate near the context window limit (no 'lost in the middle' falloff)
How to read the score
The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better.
Official methodology note: RULER 128K - long context retrieval benchmark evaluated at 128K tokens. As evaluated on the LFM3-1.7B card: model served via vLLM in BF16 at its full 262,144-token context, temperature 0.6, top_p 0.95, no output limit other than the context length.
Frequently asked questions
What is the RULER 128K benchmark?
RULER 128K is a benchmark for reading and reasoning over very long inputs. The official description says: 'RULER 128K - long context retrieval benchmark evaluated at 128K tokens'. Below we describe in simple words what benchmarks of this type generally examine.
What does RULER 128K test?
RULER 128K mainly measures reading and reasoning over very long inputs. Concretely, it checks: finding relevant facts buried in documents hundreds of thousands of tokens long; retrieving multiple needles hidden across the full context; reasoning over whole books, codebases or conversation histories.
How is the RULER 128K score calculated?
The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better. The declared maximum score of this benchmark is 100.0.
Where do I get RULER 128K data and scores?
The official benchmark page/reference is linked on this page. Every score on Open LLM Radar is also available as open data: download the scores JSON (http://open-llm-radar-api:8000/api/v1/benchmarks/ruler-128k/scores) or pull it directly from the Open LLM Radar API with any HTTP client.
Benchmark data & downloads
Where to get, inspect and reuse the data behind this benchmark.