Mandotosh/risk-routed-kv-exact-recall-benchmark

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

1. single_needle: retrieve a secret marker from a long filler context.
2. multi_needle: retrieve the marker associated with a queried label such as BLUE.
3. kv_retrieval: retrieve a key-value record such as a city access code.
4. code_string: retrieve the exact string returned by a small code block.
5. 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.

Related code

https://github.com/Ahmet2001/-risk-routed-heterogeneous-kv-memory