{ "architecture": { "class": "Spark2_5ForCausalLM", "fixture": { "context_limit": 256, "gate_attn_act_mode": "sigmoid", "head_dim": 16, "headwise_attn_output_gate": true, "hidden_act": "gelu", "hidden_size": 64, "intermediate_size": 128, "layer_types": [ "sliding_attention", "sliding_attention", "sliding_attention", "full_attention" ], "num_attention_heads": 4, "num_hidden_layers": 4, "num_key_value_heads": 1, "partial_rotary_dimensions": { "full_attention": 4, "sliding_attention": 16 }, "sliding_window": 32, "tie_word_embeddings": true, "vocab_size": 260 }, "installed_native_class": false, "model_type": "spark2_5", "scope": "Complete text causal LM; no multimodal or MTP scope" }, "commands": { "binding_note": "Provide a sealed panel built for the actual checkpoint tokenizer, not the tiny tokenizer for original weights. Every $NAME is an explicit caller-supplied shell variable; SHA fields are actual immutable Hub commits. AUTHOR is mandatory user attribution.", "publish": [ "python", "release.py", "--publish", "--out", "$NEW_STAGE_DIR" ], "stage": [ "python", "release.py", "--stage-only", "--out", "$NEW_STAGE_DIR", "--qfs-root", "$QFS" ], "strict_validate": [ "python", "$QFS/bin/fidelity_dataset.py", "validate", "$OUTPUT_DIR", "--strict", "--verify-tensors", "--json", "$STRICT_REPORT" ], "third_party_capture_argv": [ "python", "$QFS/engines/tools/hf_capture.py", "--model", "$MODEL_DIR", "--weights-repository", "$WEIGHTS_REPO", "--model-revision", "$WEIGHTS_SHA", "--trust-remote-code", "--code-repository", "malaiwah/spark2-5-tiny-random-bf16", "--code-revision", "dd4b081b0a6b042771cbfdaedf82ef345e59f5e9", "--panel", "$PANEL_DIR", "--panel-role", "final", "--panel-tokenizer-root", "$MODEL_DIR", "--role", "root", "--lane", "other", "--device", "cpu", "--dtype", "bfloat16", "--schedule", "window-outer", "--dataset-id", "$DATASET_ID", "--dataset-name", "$DATASET_NAME", "--repository", "$EVIDENCE_REPO", "--author", "$AUTHOR", "--out", "$OUTPUT_DIR" ], "verify": [ "python", "$QFS/bin/fidelity_dataset.py", "verify", "$OUTPUT_DIR", "--verify-tensors", "--json", "$VERIFY_REPORT" ] }, "licensing": { "fixture_weights": "Apache-2.0; independently randomly initialized", "reviewed_runtime": "Apache-2.0; source notices retained", "upstream_weights": "Apache-2.0", "workflow_attribution": "Inherited MIT notice retained in NOTICE" }, "original_format_limits": { "measured_original_weights": false, "not_proven": [ "production-size inference", "million-token context", "GPU/NPU kernels", "FP8/MXFP4/quantized formats", "training", "arbitrary batched/offset position_ids", "quality or accuracy", "cross-hardware identity" ], "upstream_context": 1048576, "upstream_dtype": "bfloat16", "upstream_parameters": 4112079360 }, "prerequisites": { "additional_dependencies": [], "code_consent": "--trust-remote-code required; --code-revision full40-hex immutable commit; --code-repository selects audited fork without changing weights/config bytes", "device": "cpu", "network": "Local model weights, online immutable Hub code verification permitted; not entirely offline", "python": "3.12", "requirements": "requirements-cpu.txt", "schedule": "window-outer", "threads": 2, "torch": "2.11.0+cpu", "transformers": "5.16.1" }, "schema": "malaiwah.spark2-5-support.v1", "upstream": { "repository": "XHToken/Spark-X2.5-4B", "revision": "5e10fcc0286756aebf7c41dc52c1e42d95c70281" }, "verified_runtime": { "code_repository": "malaiwah/spark2-5-tiny-random-bf16", "code_revision": "dd4b081b0a6b042771cbfdaedf82ef345e59f5e9", "comparison": { "comparator": { "accumulation_dtype": "float64", "bf16_reduced_precision_reduction": false, "cublas_workspace_config": null, "deterministic_algorithms": true, "device": "cpu", "estimator_backend": "torch:k6_kld_report._token_kld", "head_applied_candidate_tensor_content_sha256": "90323a1776a8b913101cea3fa71a8b5f66a60d6f5a73b32038a2c01f0ed3fe2d", "head_applied_reference_tensor_content_sha256": "90323a1776a8b913101cea3fa71a8b5f66a60d6f5a73b32038a2c01f0ed3fe2d", "head_applied_tensor_content_sha256": "90323a1776a8b913101cea3fa71a8b5f66a60d6f5a73b32038a2c01f0ed3fe2d", "logprob_dtype": "float64", "position_block": 128, "replay_backend": "numpy:cpu:float32", "replay_env": { "blas_configuration": "OpenBLAS 0.3.34.106.0 USE64BITINT DYNAMIC_ARCH NO_AFFINITY Haswell MAX_THREADS=64", "blas_name": "scipy-openblas", "blas_threads": 2, "blas_threads_source": "ctypes:scipy_openblas_get_num_threads64_", "blas_version": "0.3.34.106.0", "cpu_model": "Intel(R) Xeon(R) CPU X5570 @ 2.93GHz", "library": "numpy", "numpy_version": "2.5.3", "replay_dtype": "float32" }, "replay_peak_device_bytes": null, "short_circuited": true, "source_file_hashes_verified": true, "tensor_content_digests_verified": true, "tf32": false, "tool": { "entrypoint": "bin/fidelity-dataset compare", "note": "the fp64 estimator is k6_kld_report._token_kld, imported not copied, whenever torch is importable; estimator_backend records which path ran." }, "two_pass": true, "vocab_chunk": 8192 }, "comparison_kind": "reproduction_confirmation", "kl": { "max": 0.0, "mean": 0.0, "median": 0.0, "p95": 0.0, "p99": 0.0, "p99_9": 0.0 }, "measurement_scope": { "contexts": 4, "covers_full_panel": true, "position_filter": "all", "positions_per_context": 63, "scored_positions": 252, "subset_detail": null }, "metric": { "direction": "reference_to_candidate", "direction_label": "KL(reference || candidate)", "higher_is_better": false, "name": "mean_tokenwise_kld", "units": "nats", "value": 0.0 }, "note": "QFS retains its hash-proof result and top-1 value for reproduction confirmations; --force-compute independently runs the estimator and requires its tokenwise KL array to agree exactly. short_circuited=true therefore does not mean replay was skipped.", "receipt": "comparison/comparison-receipt.json", "receipt_sha256": "19048f92026f9d3643c1bc855b06b9952d206a86d31e3635c0246d747937d715", "self_compare": { "asserted_exact_zero": true, "capture_content_digest_equal": true, "expected_tokenwise_bytes": 2144, "expected_tokenwise_sha256": "896ad5b6b732dce1588636c3af3a8a93af8eb1250df72809f6e26fc29914f65d", "force_compute_agreed": true, "head_digest_equal": true, "weights_identity_equal": true }, "tokenwise": { "bytes": 2144, "path": "tokenwise-kld.npy", "sha256": "896ad5b6b732dce1588636c3af3a8a93af8eb1250df72809f6e26fc29914f65d" }, "top1_agreement": 1.0 }, "cpu_report_sha256": "9f2119f50b3d6e7b7d8dd8af8772fd373d1ce3c5b598fc33d83c21d006c2bce2", "cpu_threads": 2, "device": "cpu", "evidence_repository": "malaiwah/spark2-5-tiny-cpu-repro-v1", "model_class": "Spark2_5ForCausalLM", "native_report_sha256": "73f21cdbac7ca5e301de182342de59d2242f08f94376d3f7c18d8e9f7b202d41", "scope": "Complete text-only Spark2_5ForCausalLM; 3 sliding/1 full attention, partial RoPE dimensions 16/4, headwise sigmoid gates, GELU gated MLP, tied vocabulary head. Random BF16 only; no learned capabilities, million-token context, accelerator, training, quantized-format, or production-size evidence.", "status": "passed", "torch": "2.11.0+cpu", "transformers": "5.16.1", "weights_revision": "dd4b081b0a6b042771cbfdaedf82ef345e59f5e9" } }