Add calibrated Q4_K_M and BF16 GGUF download options

.gitattributes

@@ -37,3 +37,5 @@ Jev-Style-v2-Q8_0-Calibrated.gguf filter=lfs diff=lfs merge=lfs -text
 figures/benchmark.png filter=lfs diff=lfs merge=lfs -text
 figures/calibration.png filter=lfs diff=lfs merge=lfs -text
 figures/robustness.png filter=lfs diff=lfs merge=lfs -text
+Jev-Style-v2-Q4_K_M-Calibrated.gguf filter=lfs diff=lfs merge=lfs -text
+Jev-Style-v2-BF16-Calibrated.gguf filter=lfs diff=lfs merge=lfs -text

Jev-Style-v2-BF16-Calibrated.calibration.json

@@ -0,0 +1,56 @@
+{
+  "temperature": 1.0,
+  "fitted_temperature_folded": 1.0408715111841746,
+  "temperature_folded": true,
+  "folded_tensor": "output_norm.weight",
+  "source_calibration": {
+    "temperature": 1.0408715111841746,
+    "calibration_n": 3100,
+    "objective": "sample_mean_soft_cross_entropy",
+    "bounds": [
+      0.05,
+      20
+    ],
+    "nll_before": 0.5144028513612269,
+    "nll_after": 0.5140996476734587,
+    "backend": "gguf",
+    "model": "source_project/h100_v2/results/Jev-Style-v2-BF16-uncalibrated.gguf",
+    "source_sha256": "2fde7f45dce3440abfde145bb30ad61e2643b1f853866b5760b235685328dc1c"
+  },
+  "input_sha256": "79fb883ffc5831f165c11ca0119c7702b3c5209c916d939f2555fb1c7d45413c",
+  "output_sha256": "8baa111eec6e30a5c9e97d559b53127a19ef30171e8aed26673153b4f77adcbb",
+  "validation_required": false,
+  "validation": {
+    "format": "BF16",
+    "n": 500,
+    "argmax_agreement": 0.996,
+    "cuda_same_subset": {
+      "accuracy": 0.7910177949703642,
+      "macro_f1": 0.7760857891780776,
+      "nll": 0.6000106706289864,
+      "brier": 0.317876961372947,
+      "ece": 0.16696329399611096
+    },
+    "gguf_same_subset": {
+      "accuracy": 0.7936151975677668,
+      "macro_f1": 0.77850284511292,
+      "nll": 0.6007551256850553,
+      "brier": 0.31813150017828207,
+      "ece": 0.16639447171288832
+    },
+    "accuracy_difference": 0.0025974025974025983,
+    "nll_difference": 0.0007444550560689045,
+    "temperature_folded": true,
+    "runtime_temperature": 1.0,
+    "calibration_n": 3100,
+    "criteria": {
+      "agreement_min": 0.99,
+      "accuracy_loss_max": 0.005,
+      "nll_increase_max": 0.015
+    },
+    "calibration_backend": "llama.cpp batched, checked against native readout",
+    "evaluation_backend": "llama.cpp native single-sequence exact option logits",
+    "weight_bytes": 3775700896,
+    "passed": true
+  }
+}

Jev-Style-v2-BF16-Calibrated.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8baa111eec6e30a5c9e97d559b53127a19ef30171e8aed26673153b4f77adcbb
+size 3775700896

Jev-Style-v2-Q4_K_M-Calibrated.calibration.json

@@ -0,0 +1,56 @@
+{
+  "temperature": 1.0,
+  "fitted_temperature_folded": 1.0123069568523906,
+  "temperature_folded": true,
+  "folded_tensor": "output_norm.weight",
+  "source_calibration": {
+    "temperature": 1.0123069568523906,
+    "calibration_n": 3100,
+    "objective": "sample_mean_soft_cross_entropy",
+    "bounds": [
+      0.05,
+      20
+    ],
+    "nll_before": 0.5255198219301063,
+    "nll_after": 0.5254917547798398,
+    "backend": "gguf",
+    "model": "source_project/h100_v2/results/Jev-Style-v2-Q4_K_M.gguf",
+    "source_sha256": "2fde7f45dce3440abfde145bb30ad61e2643b1f853866b5760b235685328dc1c"
+  },
+  "input_sha256": "88982768b42a345708c93347ed49b1efebf1be366efe61c202d7ca1e931822fd",
+  "output_sha256": "c697d3b29d07fdd37b6ebeb5c98066f4c31632162adca6258db23f75184eb0c4",
+  "validation_required": false,
+  "validation": {
+    "format": "Q4_K_M",
+    "n": 500,
+    "argmax_agreement": 0.914,
+    "cuda_same_subset": {
+      "accuracy": 0.7910177949703642,
+      "macro_f1": 0.7760857891780776,
+      "nll": 0.6000106706289864,
+      "brier": 0.317876961372947,
+      "ece": 0.16696329399611096
+    },
+    "gguf_same_subset": {
+      "accuracy": 0.7818348025857907,
+      "macro_f1": 0.764930756852403,
+      "nll": 0.6474576399658714,
+      "brier": 0.3418843970562367,
+      "ece": 0.17704328656854143
+    },
+    "accuracy_difference": -0.00918299238457343,
+    "nll_difference": 0.04744696933688497,
+    "temperature_folded": true,
+    "runtime_temperature": 1.0,
+    "calibration_n": 3100,
+    "criteria": {
+      "agreement_min": 0.9,
+      "accuracy_loss_max": 0.03,
+      "nll_increase_max": 0.1
+    },
+    "calibration_backend": "llama.cpp batched, checked against native readout",
+    "evaluation_backend": "llama.cpp native single-sequence exact option logits",
+    "weight_bytes": 1274388384,
+    "passed": true
+  }
+}

Jev-Style-v2-Q4_K_M-Calibrated.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c697d3b29d07fdd37b6ebeb5c98066f4c31632162adca6258db23f75184eb0c4
+size 1274388384

README.md

@@ -14,17 +14,26 @@ tags:
 - jev-style
 - single-prefill
 ---
-# Jev-Style-Qwen3.5-2B-Decision v2 (GGUF Q8_0)
+# Jev-Style-Qwen3.5-2B-Decision v2 (GGUF)
 
 A **Jev-style decision model** for classification, routing and typed choices. Give it a state, a question and a list of options; one prefill returns a selected option **with calibrated probabilities**.
 
 | Build | Weight size | Inference |
 |---|---:|---|
 | [HF BF16](https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2) | 3.76 GB | Transformers + decision client |
-| **GGUF Q8_0 · this repository** | 2.01 GB | llama.cpp + decision client |
+| **GGUF · this repository** | 1.27–3.78 GB | Q4_K_M / Q8_0 / BF16 · llama.cpp |
 | [MLX BF16](https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2-MLX-bf16) | 3.76 GB | Apple Silicon + native MLX client |
 
-**Download this build:** [Jev-Style-v2-Q8_0-Calibrated.gguf](https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2-GGUF/resolve/main/Jev-Style-v2-Q8_0-Calibrated.gguf?download=true). The repository also includes its calibration, inference client and evaluation records.
+**Choose a precision:**
+
+| Precision | File size | Choice agreement | Macro accuracy |
+|---|---:|---:|---:|
+| [Q4_K_M](https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2-GGUF/resolve/main/Jev-Style-v2-Q4_K_M-Calibrated.gguf?download=true) | 1.27 GB | 91.4% | 78.18% |
+| [Q8_0](https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2-GGUF/resolve/main/Jev-Style-v2-Q8_0-Calibrated.gguf?download=true) | 2.01 GB | 99.2% | 78.69% |
+| [BF16](https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2-GGUF/resolve/main/Jev-Style-v2-BF16-Calibrated.gguf?download=true) | 3.78 GB | 99.6% | 79.36% |
+
+All three files include independently fitted calibration; use **runtime temperature 1.0**. Agreement is against CUDA merged BF16 on the same frozen 500-decision subset; accuracy is the task-macro average over its real-label examples. [Full precision comparison](evaluation/quantization_summary.json).
+
 
 ## Results
 
@@ -91,11 +100,15 @@ The separate typed-decisions group contains 2,000 teacher-reference decisions fr
 
 ## Deployment validation
 
+GGUF is available in **Q4_K_M, Q8_0 and BF16**, each with its own calibration and verification record.
+
 | Released format | Weight size | Validated result | Evaluation set |
 |---|---:|---|---|
 | HF BF16 | 3.76 GB | **81.27%** real-label macro accuracy | Full 3,277 real-label decisions |
 | Native MLX BF16 | 3.76 GB | **99.6%** choice agreement with CUDA BF16 | Frozen 500-decision deployment subset |
 | Calibrated GGUF Q8_0 | 2.01 GB | **99.2%** choice agreement with CUDA BF16 | Same 500-decision deployment subset |
+| Calibrated GGUF Q4_K_M | 1.27 GB | **91.4%** choice agreement with CUDA BF16 | Same 500-decision deployment subset |
+| Calibrated GGUF BF16 | 3.78 GB | **99.6%** choice agreement with CUDA BF16 | Same 500-decision deployment subset |
 
 Each deployment format has its own validation record. Native MLX packaging reproduces the verified MLX client's logits exactly on all 500 deployment cases. The Q8_0 model is approximately **46.7% smaller** than the BF16 GGUF export.
 
@@ -133,6 +146,8 @@ python jev_decision_client.py --url http://127.0.0.1:8080 \
   --options negative positive
 ```
 
+The quick-start command selects Q8_0. To use Q4_K_M or BF16, replace its model filename with `Jev-Style-v2-Q4_K_M-Calibrated.gguf` or `Jev-Style-v2-BF16-Calibrated.gguf`.
+
 Use a llama.cpp build with Qwen3.5 support. Conversion and native evaluation used commit `b29c606e28a01b1bc8c1351026a0fa6e616bf6c4`. The client uses the native `/completion` endpoint, requests complete declared-option log-probabilities and increases the candidate count as needed. The supplied `gguf_logits.cpp` reads all declared-option logits directly through the C API.
 
 **Runtime calibration temperature is 1.0** for this file: its fitted temperature has already been incorporated. The accompanying calibration JSON records the exact settings and checksum. Serve the raw decision prompt shown below, with the full declared option list.

SHA256SUMS.json

@@ -12,8 +12,8 @@
     "sha256": "50cbab8a892c5f2993b8c7351a99182507472def3b1374558308605d99b86b32"
   },
   "README.md": {
-    "bytes": 10788,
-    "sha256": "e2fc4983b02dec5266c219b9333e0ef59dcdb88a4d9393eb686a734036c73963"
+    "bytes": 11957,
+    "sha256": "37b0885c870ef69cf315d4e8bdae3e09c81972f0fd70d4417284fef50a83ddda"
   },
   "evaluation/baseline_sensitivity.json": {
     "bytes": 2015,
@@ -82,5 +82,45 @@
   "figures/calibration.png": {
     "bytes": 183790,
     "sha256": "2a4e163354b0d0d31bf5fcdfcfbcf726fb152fb2f6419432ddddaec3b475a921"
+  },
+  "evaluation/quantization_summary.json": {
+    "bytes": 5465,
+    "sha256": "45c350f225f9a403d036d96a80ef14ca9cd89bb9c25af82617071833846eefd6"
+  },
+  "evaluation/calibration_batch_parity.json": {
+    "bytes": 147,
+    "sha256": "1f819d59b9ebf6ffb79953832c3f7ae4564d7bae7fa68209b6583a9d92db2bf5"
+  },
+  "Jev-Style-v2-Q4_K_M-Calibrated.gguf": {
+    "bytes": 1274388384,
+    "sha256": "c697d3b29d07fdd37b6ebeb5c98066f4c31632162adca6258db23f75184eb0c4"
+  },
+  "Jev-Style-v2-Q4_K_M-Calibrated.calibration.json": {
+    "bytes": 1824,
+    "sha256": "4a5abb23d86ddd44b8b6b1054b48a0a9c40560cc0ee3edd24ae6be8e623ad5f1"
+  },
+  "evaluation/q4_k_m_deployment.json": {
+    "bytes": 941,
+    "sha256": "b67e128b7e46287a84ce487171882b4b8ece888c28cfa92d09e72fefec298bcc"
+  },
+  "evaluation/q4_k_m_http_smoke.json": {
+    "bytes": 694,
+    "sha256": "5a0e2b0696e3aa918aff6e8ad2ebf1532091edf81a10815bf3add221c36932d3"
+  },
+  "Jev-Style-v2-BF16-Calibrated.gguf": {
+    "bytes": 3775700896,
+    "sha256": "8baa111eec6e30a5c9e97d559b53127a19ef30171e8aed26673153b4f77adcbb"
+  },
+  "Jev-Style-v2-BF16-Calibrated.calibration.json": {
+    "bytes": 1840,
+    "sha256": "e5eb54d405d213c751ba61b334a8370d03332c915d9c49725da3b7a6ee295109"
+  },
+  "evaluation/bf16_deployment.json": {
+    "bytes": 946,
+    "sha256": "19cbc98b2c892509f804b4b758ab5009dbcd5fef9807b4cbedb4f7114a4efab2"
+  },
+  "evaluation/bf16_http_smoke.json": {
+    "bytes": 690,
+    "sha256": "d729e4dd7ab5d425e5eb031ca1dc92d74302a7971d9aeeba782c15e4904a5f59"
   }
 }

evaluation/bf16_deployment.json

@@ -0,0 +1,33 @@
+{
+  "format": "BF16",
+  "n": 500,
+  "argmax_agreement": 0.996,
+  "cuda_same_subset": {
+    "accuracy": 0.7910177949703642,
+    "macro_f1": 0.7760857891780776,
+    "nll": 0.6000106706289864,
+    "brier": 0.317876961372947,
+    "ece": 0.16696329399611096
+  },
+  "gguf_same_subset": {
+    "accuracy": 0.7936151975677668,
+    "macro_f1": 0.77850284511292,
+    "nll": 0.6007551256850553,
+    "brier": 0.31813150017828207,
+    "ece": 0.16639447171288832
+  },
+  "accuracy_difference": 0.0025974025974025983,
+  "nll_difference": 0.0007444550560689045,
+  "temperature_folded": true,
+  "runtime_temperature": 1.0,
+  "calibration_n": 3100,
+  "criteria": {
+    "agreement_min": 0.99,
+    "accuracy_loss_max": 0.005,
+    "nll_increase_max": 0.015
+  },
+  "calibration_backend": "llama.cpp batched, checked against native readout",
+  "evaluation_backend": "llama.cpp native single-sequence exact option logits",
+  "weight_bytes": 3775700896,
+  "passed": true
+}

evaluation/bf16_http_smoke.json

@@ -0,0 +1,25 @@
+{
+  "passed": true,
+  "model": "Jev-Style-v2-BF16-Calibrated.gguf",
+  "cases": [
+    {
+      "options": 2,
+      "choice": "positive",
+      "candidate_count": 64,
+      "max_probability_difference_vs_native": 5.900205922393376e-06
+    },
+    {
+      "options": 4,
+      "choice": "science and technology",
+      "candidate_count": 64,
+      "max_probability_difference_vs_native": 0.0006067566569580851
+    },
+    {
+      "options": 5,
+      "choice": "neutral",
+      "candidate_count": 64,
+      "max_probability_difference_vs_native": 0.001473201814370717
+    }
+  ],
+  "server": "version: 0.4.1 (build 10964, commit b29c606e2)\nbuilt with AppleClang 17.0.0.17000604 for Darwin arm64"
+}

evaluation/calibration_batch_parity.json

@@ -0,0 +1,7 @@
+{
+  "n": 32,
+  "max_probability_difference": 3.980138511150422e-05,
+  "argmax_agreement": 1.0,
+  "seconds": 13.387850046157837,
+  "passed": true
+}

evaluation/q4_k_m_deployment.json

@@ -0,0 +1,33 @@
+{
+  "format": "Q4_K_M",
+  "n": 500,
+  "argmax_agreement": 0.914,
+  "cuda_same_subset": {
+    "accuracy": 0.7910177949703642,
+    "macro_f1": 0.7760857891780776,
+    "nll": 0.6000106706289864,
+    "brier": 0.317876961372947,
+    "ece": 0.16696329399611096
+  },
+  "gguf_same_subset": {
+    "accuracy": 0.7818348025857907,
+    "macro_f1": 0.764930756852403,
+    "nll": 0.6474576399658714,
+    "brier": 0.3418843970562367,
+    "ece": 0.17704328656854143
+  },
+  "accuracy_difference": -0.00918299238457343,
+  "nll_difference": 0.04744696933688497,
+  "temperature_folded": true,
+  "runtime_temperature": 1.0,
+  "calibration_n": 3100,
+  "criteria": {
+    "agreement_min": 0.9,
+    "accuracy_loss_max": 0.03,
+    "nll_increase_max": 0.1
+  },
+  "calibration_backend": "llama.cpp batched, checked against native readout",
+  "evaluation_backend": "llama.cpp native single-sequence exact option logits",
+  "weight_bytes": 1274388384,
+  "passed": true
+}

evaluation/q4_k_m_http_smoke.json

@@ -0,0 +1,25 @@
+{
+  "passed": true,
+  "model": "Jev-Style-v2-Q4_K_M-Calibrated.gguf",
+  "cases": [
+    {
+      "options": 2,
+      "choice": "positive",
+      "candidate_count": 64,
+      "max_probability_difference_vs_native": 8.535655050662637e-07
+    },
+    {
+      "options": 4,
+      "choice": "science and technology",
+      "candidate_count": 64,
+      "max_probability_difference_vs_native": 1.7518853354103747e-05
+    },
+    {
+      "options": 5,
+      "choice": "neutral",
+      "candidate_count": 64,
+      "max_probability_difference_vs_native": 9.285758503529973e-05
+    }
+  ],
+  "server": "version: 0.4.1 (build 10964, commit b29c606e2)\nbuilt with AppleClang 17.0.0.17000604 for Darwin arm64"
+}

evaluation/quantization_summary.json

@@ -0,0 +1,150 @@
+{
+  "evaluation_decisions": 500,
+  "comparison": "same frozen subset vs CUDA merged BF16",
+  "metric": "real-label task-macro accuracy; teacher-reference decisions excluded from this macro",
+  "variants": {
+    "Q4_K_M": {
+      "filename": "Jev-Style-v2-Q4_K_M-Calibrated.gguf",
+      "weight_bytes": 1274388384,
+      "sha256": "c697d3b29d07fdd37b6ebeb5c98066f4c31632162adca6258db23f75184eb0c4",
+      "argmax_agreement": 0.914,
+      "real_label_macro_accuracy": 0.7818348025857907,
+      "real_label_macro": {
+        "accuracy": 0.7818348025857907,
+        "macro_f1": 0.764930756852403,
+        "nll": 0.6474576399658714,
+        "brier": 0.3418843970562367,
+        "ece": 0.17704328656854143
+      },
+      "calibration_n": 3100,
+      "temperature_folded": true,
+      "runtime_temperature": 1.0,
+      "validation": {
+        "format": "Q4_K_M",
+        "n": 500,
+        "argmax_agreement": 0.914,
+        "cuda_same_subset": {
+          "accuracy": 0.7910177949703642,
+          "macro_f1": 0.7760857891780776,
+          "nll": 0.6000106706289864,
+          "brier": 0.317876961372947,
+          "ece": 0.16696329399611096
+        },
+        "gguf_same_subset": {
+          "accuracy": 0.7818348025857907,
+          "macro_f1": 0.764930756852403,
+          "nll": 0.6474576399658714,
+          "brier": 0.3418843970562367,
+          "ece": 0.17704328656854143
+        },
+        "accuracy_difference": -0.00918299238457343,
+        "nll_difference": 0.04744696933688497,
+        "temperature_folded": true,
+        "runtime_temperature": 1.0,
+        "calibration_n": 3100,
+        "criteria": {
+          "agreement_min": 0.9,
+          "accuracy_loss_max": 0.03,
+          "nll_increase_max": 0.1
+        },
+        "calibration_backend": "llama.cpp batched, checked against native readout",
+        "evaluation_backend": "llama.cpp native single-sequence exact option logits",
+        "weight_bytes": 1274388384,
+        "passed": true
+      }
+    },
+    "Q8_0": {
+      "filename": "Jev-Style-v2-Q8_0-Calibrated.gguf",
+      "weight_bytes": 2012004256,
+      "sha256": "5c2aa0d35b24a27f03228b2c62ebaaebd9b5b785844634d4217278d822751494",
+      "argmax_agreement": 0.992,
+      "real_label_macro_accuracy": 0.7868855635654054,
+      "real_label_macro": {
+        "accuracy": 0.7868855635654054,
+        "macro_f1": 0.773615445189482,
+        "nll": 0.6015417570028033,
+        "brier": 0.318930434743987,
+        "ece": 0.16630794459650552
+      },
+      "calibration_n": 3100,
+      "temperature_folded": true,
+      "runtime_temperature": 1.0,
+      "validation": {
+        "n": 500,
+        "argmax_agreement": 0.992,
+        "cuda_same_subset": {
+          "accuracy": 0.7910177949703642,
+          "macro_f1": 0.7760857891780776,
+          "nll": 0.6000106706289864,
+          "brier": 0.317876961372947,
+          "ece": 0.16696329399611096
+        },
+        "q8_same_subset": {
+          "accuracy": 0.7868855635654054,
+          "macro_f1": 0.773615445189482,
+          "nll": 0.6015417570028033,
+          "brier": 0.318930434743987,
+          "ece": 0.16630794459650552
+        },
+        "accuracy_difference": -0.004132231404958775,
+        "nll_difference": 0.0015310863738169367,
+        "temperature_folded": true,
+        "passed": true,
+        "practical_deployment_gate_passed": true,
+        "initial_strict_accuracy_target_met_on_subset": false,
+        "initial_accuracy_loss_target": 0.003,
+        "note": "500-example subset check: 99.2% agreement. Observed task-macro loss is 0.413 pp, so the initial 0.3 pp accuracy goal is not met on this subset. Prefer BF16/MLX when accuracy is the priority; this is not a bound on population degradation."
+      }
+    },
+    "BF16": {
+      "filename": "Jev-Style-v2-BF16-Calibrated.gguf",
+      "weight_bytes": 3775700896,
+      "sha256": "8baa111eec6e30a5c9e97d559b53127a19ef30171e8aed26673153b4f77adcbb",
+      "argmax_agreement": 0.996,
+      "real_label_macro_accuracy": 0.7936151975677668,
+      "real_label_macro": {
+        "accuracy": 0.7936151975677668,
+        "macro_f1": 0.77850284511292,
+        "nll": 0.6007551256850553,
+        "brier": 0.31813150017828207,
+        "ece": 0.16639447171288832
+      },
+      "calibration_n": 3100,
+      "temperature_folded": true,
+      "runtime_temperature": 1.0,
+      "validation": {
+        "format": "BF16",
+        "n": 500,
+        "argmax_agreement": 0.996,
+        "cuda_same_subset": {
+          "accuracy": 0.7910177949703642,
+          "macro_f1": 0.7760857891780776,
+          "nll": 0.6000106706289864,
+          "brier": 0.317876961372947,
+          "ece": 0.16696329399611096
+        },
+        "gguf_same_subset": {
+          "accuracy": 0.7936151975677668,
+          "macro_f1": 0.77850284511292,
+          "nll": 0.6007551256850553,
+          "brier": 0.31813150017828207,
+          "ece": 0.16639447171288832
+        },
+        "accuracy_difference": 0.0025974025974025983,
+        "nll_difference": 0.0007444550560689045,
+        "temperature_folded": true,
+        "runtime_temperature": 1.0,
+        "calibration_n": 3100,
+        "criteria": {
+          "agreement_min": 0.99,
+          "accuracy_loss_max": 0.005,
+          "nll_increase_max": 0.015
+        },
+        "calibration_backend": "llama.cpp batched, checked against native readout",
+        "evaluation_backend": "llama.cpp native single-sequence exact option logits",
+        "weight_bytes": 3775700896,
+        "passed": true
+      }
+    }
+  }
+}