Datasets:

Mandotosh
/

risk-routed-kv-exact-recall-benchmark

license: mit
language:
  - en
task_categories:
  - text-generation
  - question-answering
pretty_name: Risk-Routed KV Exact-Recall Benchmark
size_categories:
  - n<1K

Risk-Routed KV Exact-Recall Benchmark

This dataset contains controlled synthetic exact-recall examples used to evaluate risk-routed heterogeneous KV memory policies for long-context Transformer inference.

The benchmark is designed for testing whether a model can retrieve exact strings from long contexts under different KV-cache policies:

Full KV
Uniform low-bit Quantized KV
Risk-routed heterogeneous KV, where exact-critical spans stay in Full KV and background context is quantized

Tasks

single_needle: retrieve a secret marker from a long filler context.
multi_needle: retrieve the marker associated with a queried label such as BLUE.
kv_retrieval: retrieve a key-value record such as a city access code.
code_string: retrieve the exact string returned by a small code block.
date_negation: retrieve the approved final date while ignoring a negated distractor.

Schema

Each JSONL row has:

{
  "id": "single_needle-8192w-0",
  "task": "single_needle",
  "context_words_target": 8192,
  "context": "...long context...",
  "query": "Which marker was labeled as the secret marker?",
  "gold": "MK53-4897X",
  "distractors": ["ZX17-2044Q", "ALPHA-77K2"],
  "choices": ["MK53-4897X", "ZX17-2044Q", "ALPHA-77K2"],
  "exact_spans": ["MK53-4897X"],
  "source": "synthetic_controlled_exact_recall"
}

Intended evaluation

For forced-choice evaluation, compute answer sequence NLL for each candidate in choices after the context and query. The prediction is correct if the gold answer has the lowest NLL.

Risk-Routed KV Exact-Recall Benchmark

Tasks

Schema

Intended evaluation

Related code