{
  "name": "olmoe-1b-7b-compacted-5b",
  "version": "1.0.0",
  "description": "Cross-architecture validation artifact for the \u00a74.1.3.4 calibration-aware expert importance methodology. OLMoE-1B-7B-0924-Instruct (the smallest serious MoE on HuggingFace, fully-open Allen AI release) compacted from 64 experts per layer to 48 via per-layer-normalized activation-count importance ranking on a held-out Python code calibration corpus. Hardware-measured 36.0 HumanEval / 31.7 HumanEval+ vs the unmodified base's 40.9 / 36.6 \u2014 within \u22124.9 / \u22124.9 of the base anchor. The negative-baseline broad-corpus variant scored 28.0 / 26.2 (\u0394 \u221212.9 / \u221210.4); the +8.0 / +5.5 swing from changing only the calibration corpus is the second empirical anchor for \u00a74.1.3.4 (the first was Qwen3-Coder-30B-A3B with a +9.7 swing). Two architectures (`Qwen3MoeForCausalLM` and `OlmoeForCausalLM`) now empirically validate the cross-architecture invariance claim: the metric is architecture-invariant, the calibration-corpus alignment is the lever.",
  "userSummary": "**Cross-architecture validation artifact for \u00a74.1.3.4.** OLMoE-1B-7B (the smallest serious MoE on HF, fully-open Allen AI release) compacted from 7B to ~5B via calibration-aware MoE expert pruning on a held-out Python code corpus. Hardware-measured **36.0 HumanEval** against unmodified base **40.9** (\u0394 \u22124.9, both Q5_K_M on the same 5090). The forge methodology that produced [`qwen3-coder-30b-a3b-compacted-19b-256k`](https://huggingface.co/continuum-ai/qwen3-coder-30b-a3b-compacted-19b-256k) ports to a structurally distinct MoE family (`OlmoeForCausalLM` vs `Qwen3MoeForCausalLM`) **without any modification to the forge scripts**. The negative-baseline broad-corpus variant scored 28.0 \u2014 the +8.0 swing from changing only the calibration corpus is the lever \u00a74.1.3.4 names. This is a methodology proof point, not a tier-leading artifact; OLMoE is general-purpose, not coder-specific, so HumanEval is not its strength. Use the qwen3-coder-30b-a3b artifact if you need a fits-12-GB code model.",
  "author": "continuum-ai",
  "tags": [
    "allenai",
    "olmoe",
    "moe",
    "mixture-of-experts",
    "expert-pruning",
    "calibration-aware-pruning",
    "fully-open",
    "1b-active",
    "5b",
    "gguf",
    "q5_k_m",
    "llama-cpp",
    "ollama",
    "lm-studio"
  ],
  "license": "apache-2.0",
  "source": {
    "baseModel": "allenai/OLMoE-1B-7B-0924-Instruct",
    "architecture": "olmoe",
    "isMoE": true,
    "totalParamsB": 6.92,
    "activeParamsB": 1.3,
    "numLayers": 16,
    "numExpertsPerLayer": 64,
    "numActivatedExperts": 8,
    "license": "apache-2.0"
  },
  "stages": [
    {
      "type": "expert-activation-profile",
      "calibrationCorpus": "heldout_code_python_300ex_138ktok",
      "calibrationExamples": 300,
      "calibrationTokens": 138816,
      "metricVersion": "v1.activation_count",
      "implementation": "scripts/expert_activation_profile.py",
      "notes": "Same script unchanged from the Qwen3-Coder-30B-A3B forge \u2014 first cross-architecture validation that the activation-count importance metric ports across MoE families. The hooks register on `model.layers.{L}.mlp.gate` for both Qwen3MoE and OlmoeForCausalLM (same module path)."
    },
    {
      "type": "expert-prune",
      "strategy": "calibration-aware-activation-count",
      "metric": "activation_count",
      "metricSource": "expert_activation_profile.py against heldout_code_python_300ex",
      "keepExpertsPerLayer": 48,
      "originalExpertsPerLayer": 64,
      "prunePct": 25.0,
      "expertsDropped": 768,
      "expertsRenamed": 2304,
      "routerSlicedLayers": 16,
      "perLayerNormalized": true,
      "implementation": "scripts/cpu_expert_prune_v2.py --importance-json",
      "notes": "Same script unchanged. Identical regex layout (unfused per-expert tensors at `model.layers.{L}.mlp.experts.{K}.{gate,up,down}_proj.weight`). Cross-arch portability confirmed: OlmoeForCausalLM and Qwen3MoeForCausalLM share the same prunable-unit module structure, so the script works without modification."
    },
    {
      "type": "quant",
      "format": "gguf",
      "quantTypes": [
        "Q5_K_M"
      ],
      "tool": "llama.cpp llama-quantize",
      "tiers": {
        "Q5_K_M": {
          "sizeGb": 3.6,
          "fitsVramGb": 6,
          "headlineTier": "anywhere \u2014 4 GB phones, edge devices, embedded"
        }
      }
    },
    {
      "type": "eval",
      "benchmarks": [
        {
          "name": "humaneval",
          "metric": "pass@1",
          "calibrated": true
        },
        {
          "name": "humaneval_plus",
          "metric": "pass@1",
          "calibrated": true
        }
      ],
      "calibrationAnchor": {
        "model": "allenai/OLMoE-1B-7B-0924-Instruct",
        "publishedScore": null,
        "publishedSource": "Allen AI did not publish HumanEval for OLMoE; comparison is hardware-measured against the unmodified base in the same eval pipeline",
        "measuredScore": 40.9,
        "measuredOn": "RTX 5090 32 GB \u00b7 llama.cpp llama-server \u00b7 Q5_K_M \u00b7 greedy bs=1 \u00b7 ctx 4096"
      },
      "notes": "Self-anchor calibration. HumanEval is not OLMoE's natural benchmark \u2014 OLMoE is general-purpose, not coder-specific. The 40.9 base / 36.0 student numbers are methodology validation, not tier-leading absolute quality. The artifact's value is the structural finding (cross-architecture portability + +8.0 swing from calibration alignment), not the absolute number."
    }
  ],
  "cycles": 1,
  "hardware": {
    "minVramGb": 4,
    "recommendedVramGb": 6,
    "deviceTargets": [
      "edge",
      "phone",
      "raspberry-pi",
      "rtx3050",
      "macbook-air-8gb",
      "macbook-pro-m-series"
    ]
  },
  "results": {
    "baselinePerplexity": null,
    "finalPerplexity": null,
    "improvementPct": null,
    "forgedParamsB": 5.31,
    "activeParamsB": 1.0,
    "benchmarks": [
      {
        "name": "humaneval",
        "metric": "pass@1",
        "score": 36.0,
        "baseScore": 40.9,
        "delta": -4.9,
        "calibrated": true,
        "samplesPath": "eval/humaneval/student_samples.jsonl",
        "baseSamplesPath": "eval/humaneval/base_samples.jsonl",
        "resultHash": "sha256:320dcfb889cb589ac6dcaef2c7f6481cd0ace0c5959fbea9e3255a4cc9de3ea9",
        "baseResultHash": "sha256:a56696de6f2e5cbfeeb11bfc03aced5c57ceeee39621db68ced1990ffb833050"
      },
      {
        "name": "humaneval_plus",
        "metric": "pass@1",
        "score": 31.7,
        "baseScore": 36.6,
        "delta": -4.9,
        "calibrated": true,
        "samplesPath": "eval/humaneval/student_samples.jsonl",
        "baseSamplesPath": "eval/humaneval/base_samples.jsonl",
        "resultHash": "sha256:320dcfb889cb589ac6dcaef2c7f6481cd0ace0c5959fbea9e3255a4cc9de3ea9",
        "baseResultHash": "sha256:a56696de6f2e5cbfeeb11bfc03aced5c57ceeee39621db68ced1990ffb833050"
      }
    ],
    "hardwareVerified": [
      {
        "device": "NVIDIA GeForce RTX 5090",
        "vramGb": 32,
        "format": "Q5_K_M",
        "sizeGb": 3.6,
        "tokensPerSec": null
      }
    ],
    "integrity": {
      "trustLevel": "self-attested"
    }
  },
  "priorMetricBaselines": [
    {
      "id": "olmoe-broad-corpus-2026-04-08",
      "metric": {
        "name": "calibration_aware_activation_count",
        "description": "Per-layer activation count from a BROAD calibration corpus (1/6 code, 5/6 math/knowledge/reasoning/instruction/general). Same metric as the current prune, but the calibration corpus did NOT match the eval task (HumanEval, 100% code).",
        "calibrationCorpus": "heldout_broad_300ex_50ktok",
        "calibrationCorpusSha256": null
      },
      "prune": {
        "method": "per_layer_top_k",
        "k": 48,
        "totalExpertsBefore": 64,
        "totalExpertsAfter": 48,
        "implementation": "scripts/cpu_expert_prune_v2.py --importance-json (broad calibration JSON)"
      },
      "evaluation": {
        "harness": "llama.cpp llama-server + evalplus openai backend",
        "quantization": "Q5_K_M",
        "hardware": "NVIDIA GeForce RTX 5090",
        "anchorBaseModel": "allenai/OLMoE-1B-7B-0924-Instruct",
        "anchorReproduction": {
          "humaneval": 40.9,
          "humaneval_plus": 36.6
        },
        "results": {
          "humaneval": 28.0,
          "humaneval_plus": 26.2,
          "deltaVsAnchorHumaneval": -12.9,
          "deltaVsAnchorHumanevalPlus": -10.4
        },
        "samplesPath": "eval/humaneval/student_samples_broad_calibration.jsonl"
      },
      "outcome": "negative_baseline",
      "supersededBy": "olmoe-code-corpus-2026-04-08",
      "methodologyAnchor": "https://github.com/CambrianTech/continuum/blob/main/docs/papers/PLASTICITY-COMPACTION.md#4134",
      "lessonAnchor": "Calibration corpus must match the evaluation workload. The +8.0 HumanEval swing between broad and code calibration on the same architecture, same prune budget, same hardware demonstrates that calibration-task alignment is the lever \u00a74.1.3.4 names."
    }
  ],
  "limitations": [
    "**HumanEval is not OLMoE's natural benchmark.** OLMoE is general-purpose (Allen AI), not coder-specific. The 40.9 base / 36.0 student numbers are methodology validation, not tier-leading absolute quality. For a tier-leading code model, see [`qwen3-coder-30b-a3b-compacted-19b-256k`](https://huggingface.co/continuum-ai/qwen3-coder-30b-a3b-compacted-19b-256k).",
    "**Validates \u00a74.1.3.4 cross-architecture; does NOT compete on absolute numbers.** This is the second empirical anchor for the methodology paper, alongside the Qwen3-Coder-30B-A3B v1. Together they demonstrate that the activation-count importance metric is architecture-invariant across two structurally distinct MoE families.",
    "Calibration corpus was 300 Python code examples. For non-code workloads (math/reasoning/general), the methodology will preserve OLMoE's general capability if profiled on a matching corpus \u2014 but that's a separate forge run.",
    "Single GGUF tier shipped (Q5_K_M, 3.6 GB). Q4_K_M and Q8_0 will be added in v1.1 if there's demand."
  ],
  "methodologyPaperUrl": "https://github.com/CambrianTech/continuum/blob/main/docs/papers/PLASTICITY-COMPACTION.md",
  "receipt": {
    "publications": [
      {
        "target": "huggingface",
        "url": "https://huggingface.co/continuum-ai/olmoe-1b-7b-compacted-5b",
        "publishedAt": "2026-04-08T16:36:55.037319+00:00"
      }
    ],
    "issuedAt": "2026-04-08T16:36:55.037319+00:00"
  }
}