Add MBE evaluation manifest, dataset card, and proof-of-concept result card

README.md

@@ -0,0 +1,54 @@
+---
+license: cc-by-4.0
+language:
+  - en
+tags:
+  - kv-cache
+  - llm-inference
+  - efficiency
+  - evaluation-protocol
+  - long-context
+  - benchmark
+task_categories:
+  - text-generation
+pretty_name: Matched-Budget Evaluation (MBE) for KV Cache Compression
+size_categories:
+  - n<1K
+configs:
+  - config_name: manifest
+    data_files: mbe_manifest.json
+  - config_name: results
+    data_files: cards/*.json
+---
+
+# Matched-Budget Evaluation (MBE) — KV Cache Compression
+
+A **standardized reporting protocol** for KV cache compression in LLM inference. MBE is
+not a new task benchmark; it is a thin reporting layer that fixes *which* models, tasks,
+and budgets results are reported at, so that numbers from different papers become
+comparable.
+
+- **Manifest** (`mbe_manifest.json`): the frozen evaluation specification — model suite,
+  task suite (consuming existing benchmarks: LongBench, RULER, SCBench, GSM8K, IFEval),
+  the fixed KV-budget ladder (50 / 25 / 12.5 / 6.25 %), and the required system metrics.
+  Evaluate at these exact settings so results line up.
+- **Results** (`cards/*.json`): submitted **KV Compression Cards** — one method × one
+  model, produced by the open harness under matched budgets.
+
+## Why
+Published KV cache compression results are not comparable (different models, budgets,
+tasks, system metrics). MBE fixes the axes. See the companion survey and harness:
+- Harness / protocol: https://github.com/rohithreddybc/mbe-protocol
+- Survey: "Breaking the Memory Wall: A Survey of Key-Value (KV) Cache Compression for
+  Efficient Large Language Model (LLM) Inference" (Artificial Intelligence Review, under
+  review).
+
+## How to contribute a result
+Run the harness (`run_mbe.py`) on the manifest's model + budget ladder, then submit your
+card JSON via PR to the GitHub repo or as a dataset PR here.
+
+## Citation
+See `CITATION.cff` in the GitHub repository.
+
+## License
+CC-BY-4.0. The manifest references third-party benchmarks under their own licenses.

cards/smoke_quant_qwen2.5-0.5b.json

@@ -0,0 +1,28 @@
+{
+  "harness_version": "mbe-0.1.0",
+  "run_type": "CPU smoke test (proof-of-concept, NOT the 7-8B suite)",
+  "model": "Qwen/Qwen2.5-0.5B-Instruct",
+  "model_size": "0.5B (GQA)",
+  "task": "passkey retrieval (synthetic, single needle)",
+  "n_samples": 10,
+  "hardware": "CPU / float32",
+  "method_family": "KV quantization (KIVI-style: per-channel keys, per-token values)",
+  "results": {
+    "full-FP": {
+      "kv_budget": "100%",
+      "passkey_acc": 1.0
+    },
+    "KV-8bit": {
+      "kv_budget": "50.0%",
+      "passkey_acc": 1.0
+    },
+    "KV-4bit": {
+      "kv_budget": "25.0%",
+      "passkey_acc": 1.0
+    },
+    "KV-2bit": {
+      "kv_budget": "12.5%",
+      "passkey_acc": 0.0
+    }
+  }
+}

mbe_manifest.json

@@ -0,0 +1,28 @@
+{
+  "protocol": "Matched-Budget Evaluation (MBE)",
+  "version": "0.1.0",
+  "description": "Frozen evaluation specification for KV cache compression. Report at these exact settings so results are comparable.",
+  "budget_ladder_percent": [50, 25, 12.5, 6.25],
+  "budget_definition": "Fraction of the full-cache footprint retained, computed from M_KV = 2*B*T*L*H_kv*D_head*P.",
+  "model_suite": [
+    {"id": "meta-llama/Llama-3.1-8B-Instruct", "attention": "GQA", "tier": "7-8B"},
+    {"id": "Qwen/Qwen2.5-7B-Instruct", "attention": "GQA", "tier": "7-14B"},
+    {"id": "meta-llama/Llama-3.1-70B-Instruct", "attention": "GQA", "tier": ">=70B"}
+  ],
+  "task_suite": [
+    {"name": "long_doc_qa", "source": "LongBench / SCBench", "measures": "retrieval"},
+    {"name": "multi_hop", "source": "RULER", "measures": "tracing + aggregation"},
+    {"name": "instruction_following", "source": "IFEval", "measures": "multi-instruction"},
+    {"name": "reasoning", "source": "GSM8K", "measures": "chain-of-thought (compression-sensitive)"},
+    {"name": "agentic", "source": "one long-horizon / multi-turn trace", "measures": "tool-use stability"}
+  ],
+  "required_system_metrics": [
+    "peak_kv_memory_gb",
+    "decode_throughput_tok_s",
+    "ttft_ms",
+    "max_batch_before_oom",
+    "hardware_tier"
+  ],
+  "method_metadata": ["family", "deployment_prerequisite", "composability"],
+  "notes": "MBE consumes existing public benchmarks under their own licenses; it does not redistribute them. It fixes the axes (budgets, models, metrics) and the reporting card."
+}