Update JSONLite descendant training and current evaluation metrics

PACKAGING_VERIFICATION.json

@@ -18,5 +18,5 @@
     "scope": "Packaging smoke only. A plausible but repetitive continuation is not an accuracy measurement.",
     "matches_historical_prefix": true
   },
-  "documentation_update": "Main JSONLite descendant includes additional cardinality/replay distillation SFT; base weights, tokenizer, runner and bundle unchanged."
+  "documentation_update": "Latest JSONLite joint extraction SFT and matched metrics; base weights, tokenizer, runner and bundle unchanged."
 }

README.md

@@ -175,20 +175,8 @@ not establish blanket underlying-content redistribution clearance.
 
 ## Native JSON descendant update
 
-The current [OpenSML-JSONLite-175M-Instruct](https://huggingface.co/OpenSML/OpenSML-JSONLite-175M)
-continues this pretrained base through grounded QA, requested-field and focused extraction,
-a 15,000-presentation native JSON foundation, and an additional focused extraction SFT.
-The added stage combines original JSON TRAIN replay, matched zero/one/two/three-value
-examples, and training-only frozen-parent distillation. The main model generates the
-entire JSON object itself from source text and a supplied schema, with no assembly or repair.
-
-On the same previously observed 1,444-request native JSON evaluation, complete-request
-correctness is **1,199/1,444 (83.03%)**, compared with 1,191/1,444 for the prior native release.
-Field accuracy is 91.36%, value F1 90.40%, missing-field accuracy 92.79%, and schema validity
-99.79%. The canonical subset is 1,163/1,388 (83.79%); observed playground correctness is
-44/54. These are descendant results, not scores of this pretrained base. Some domains and
-value precision/recall regress slightly, and strict automatic retention gates were not met.
-See the [JSONLite technical report](https://huggingface.co/OpenSML/OpenSML-JSONLite-175M/blob/main/TECHNICAL_REPORT.md)
-for full training provenance, matched comparison and limitations.
-
-The base weights, tokenizer, inference runner and source bundle are unchanged by this documentation update.
+The current [OpenSML-JSONLite-175M-Instruct](https://huggingface.co/OpenSML/OpenSML-JSONLite-175M) adds a joint full-weight JSON extraction SFT to its earlier grounded-QA, extraction and native JSON ancestry. The latest stage mixes original source/schema/JSON TRAIN replay, multiple-value examples and authored prose. The released main checkpoint inherits 19,200 presentations from a 21,144-presentation experiment. It directly generates the complete JSON object; validation follows without assembly or repair.
+
+On the same previously observed 1,444-request SNIPS-derived evaluation, the descendant achieves **1,271/1,444 (88.02%) complete correctness**, **94.17% field accuracy**, **92.88% extracted-value F1**, **92.42% precision**, **93.35% recall**, **95.22% missing-field accuracy** and **100% schema validity**. The observed 60-case playground scores 49/60. Compared with the previous release, 96 requests improve and 24 regress. These are descendant results on repeatedly observed research panels, not scores of this base or a fresh blind benchmark. See the [current JSONLite technical report](https://huggingface.co/OpenSML/OpenSML-JSONLite-175M/blob/main/TECHNICAL_REPORT.md) for the full provenance, protocol and limitations.
+
+The pretrained base weights, tokenizer, inference runner and source bundle are unchanged.

RUN_GUIDE.md

@@ -60,20 +60,8 @@ following or extraction. See the model card and report for measured scope.
 
 ## Native JSON descendant update
 
-The current [OpenSML-JSONLite-175M-Instruct](https://huggingface.co/OpenSML/OpenSML-JSONLite-175M)
-continues this pretrained base through grounded QA, requested-field and focused extraction,
-a 15,000-presentation native JSON foundation, and an additional focused extraction SFT.
-The added stage combines original JSON TRAIN replay, matched zero/one/two/three-value
-examples, and training-only frozen-parent distillation. The main model generates the
-entire JSON object itself from source text and a supplied schema, with no assembly or repair.
-
-On the same previously observed 1,444-request native JSON evaluation, complete-request
-correctness is **1,199/1,444 (83.03%)**, compared with 1,191/1,444 for the prior native release.
-Field accuracy is 91.36%, value F1 90.40%, missing-field accuracy 92.79%, and schema validity
-99.79%. The canonical subset is 1,163/1,388 (83.79%); observed playground correctness is
-44/54. These are descendant results, not scores of this pretrained base. Some domains and
-value precision/recall regress slightly, and strict automatic retention gates were not met.
-See the [JSONLite technical report](https://huggingface.co/OpenSML/OpenSML-JSONLite-175M/blob/main/TECHNICAL_REPORT.md)
-for full training provenance, matched comparison and limitations.
-
-The base weights, tokenizer, inference runner and source bundle are unchanged by this documentation update.
+The current [OpenSML-JSONLite-175M-Instruct](https://huggingface.co/OpenSML/OpenSML-JSONLite-175M) adds a joint full-weight JSON extraction SFT to its earlier grounded-QA, extraction and native JSON ancestry. The latest stage mixes original source/schema/JSON TRAIN replay, multiple-value examples and authored prose. The released main checkpoint inherits 19,200 presentations from a 21,144-presentation experiment. It directly generates the complete JSON object; validation follows without assembly or repair.
+
+On the same previously observed 1,444-request SNIPS-derived evaluation, the descendant achieves **1,271/1,444 (88.02%) complete correctness**, **94.17% field accuracy**, **92.88% extracted-value F1**, **92.42% precision**, **93.35% recall**, **95.22% missing-field accuracy** and **100% schema validity**. The observed 60-case playground scores 49/60. Compared with the previous release, 96 requests improve and 24 regress. These are descendant results on repeatedly observed research panels, not scores of this base or a fresh blind benchmark. See the [current JSONLite technical report](https://huggingface.co/OpenSML/OpenSML-JSONLite-175M/blob/main/TECHNICAL_REPORT.md) for the full provenance, protocol and limitations.
+
+The pretrained base weights, tokenizer, inference runner and source bundle are unchanged.

SHA256SUMS

@@ -1,12 +1,12 @@
 59ee3f5b0b4e357e34a390a72febcc13eac5493fb4a4c2d7ecdf6ba5be324ff2  CITATION.cff
 9c6d4c439acf90f9187ba6cf11f070afb3ea1515d488b52c381adf9356a1d613  EVALUATION_RESULTS.json
 c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4  LICENSE
-27478e40538a7777e0e1f8aed83f16afa6d9cf47084f6bfd39c7f8a9df515fb9  PACKAGING_VERIFICATION.json
+aa0e0453672382a7a0d7ec123d24ad83edde6d19909f141e8305235454d0d027  PACKAGING_VERIFICATION.json
 14819fc05509546bb8218153cf8494c6a3c4e2e4744ce504bdc21ef0d7d1ed59  PRETRAINING_SAMPLES.json
-f97034ba3ecbbb05f200a48cb9813ea5eb9c9e46c19f04496db14b03251a3967  README.md
-52863aa0d0b0fc25cad46c07830c244b9fa576f38b29fe39d2d086bbf8e78fae  RUN_GUIDE.md
+2e008cfeba481b8f9b37de95df231665e04b72ecc44bf04447580a295fa10d5a  README.md
+b556a9c2a63f43c904af7460936cff4e6da9ece27d2877ddb8c5eac2346b7f2a  RUN_GUIDE.md
 c3bd058575a0313f716dabb8de8a3d7b729ff413dae5f74b06525c0737dcfac3  SOURCE_FILES.json
-3ff056103137004972aaa69e1f50783b70534d951a905579cb72f151a4170d44  TECHNICAL_REPORT.md
+c76d8a9799f36f917c463dadd9d77321241f28a99bd094702887ef2700f39bb9  TECHNICAL_REPORT.md
 3235c99f09bfab6e5ee95e64a382a4975978ec7f8988d33a71d4aaeb0215bf15  THIRD_PARTY_NOTICES.md
 15a0e3b511b97503a611c0da838a7fde1c06ebed34c7103f7cda1782cc6378a4  checkpoint_metadata.json
 22eaa19ba2346558711ff5245b4a1e581a431b7b1c5df3c2342f52407ebbc74a  checkpoint_provenance.json

TECHNICAL_REPORT.md

@@ -153,20 +153,8 @@ underlying-source redistribution clearance.
 
 ## Native JSON descendant update
 
-The current [OpenSML-JSONLite-175M-Instruct](https://huggingface.co/OpenSML/OpenSML-JSONLite-175M)
-continues this pretrained base through grounded QA, requested-field and focused extraction,
-a 15,000-presentation native JSON foundation, and an additional focused extraction SFT.
-The added stage combines original JSON TRAIN replay, matched zero/one/two/three-value
-examples, and training-only frozen-parent distillation. The main model generates the
-entire JSON object itself from source text and a supplied schema, with no assembly or repair.
-
-On the same previously observed 1,444-request native JSON evaluation, complete-request
-correctness is **1,199/1,444 (83.03%)**, compared with 1,191/1,444 for the prior native release.
-Field accuracy is 91.36%, value F1 90.40%, missing-field accuracy 92.79%, and schema validity
-99.79%. The canonical subset is 1,163/1,388 (83.79%); observed playground correctness is
-44/54. These are descendant results, not scores of this pretrained base. Some domains and
-value precision/recall regress slightly, and strict automatic retention gates were not met.
-See the [JSONLite technical report](https://huggingface.co/OpenSML/OpenSML-JSONLite-175M/blob/main/TECHNICAL_REPORT.md)
-for full training provenance, matched comparison and limitations.
-
-The base weights, tokenizer, inference runner and source bundle are unchanged by this documentation update.
+The current [OpenSML-JSONLite-175M-Instruct](https://huggingface.co/OpenSML/OpenSML-JSONLite-175M) adds a joint full-weight JSON extraction SFT to its earlier grounded-QA, extraction and native JSON ancestry. The latest stage mixes original source/schema/JSON TRAIN replay, multiple-value examples and authored prose. The released main checkpoint inherits 19,200 presentations from a 21,144-presentation experiment. It directly generates the complete JSON object; validation follows without assembly or repair.
+
+On the same previously observed 1,444-request SNIPS-derived evaluation, the descendant achieves **1,271/1,444 (88.02%) complete correctness**, **94.17% field accuracy**, **92.88% extracted-value F1**, **92.42% precision**, **93.35% recall**, **95.22% missing-field accuracy** and **100% schema validity**. The observed 60-case playground scores 49/60. Compared with the previous release, 96 requests improve and 24 regress. These are descendant results on repeatedly observed research panels, not scores of this base or a fresh blind benchmark. See the [current JSONLite technical report](https://huggingface.co/OpenSML/OpenSML-JSONLite-175M/blob/main/TECHNICAL_REPORT.md) for the full provenance, protocol and limitations.
+
+The pretrained base weights, tokenizer, inference runner and source bundle are unchanged.