temsa's picture
Publish ContextPII rc4 with decoder hardening and updated evals
3460734 verified
{
"training_mix": "data/irish_core_globalpointer_focus_extended_v1",
"training_mix_meta": {
"focus_dir": "data/irish_core_globalpointer_focus_chatbot_v3",
"other_dir": "data/irish_other_pii_v1_mix",
"demo_dir": "data/globalpointer_demographic_patch_v1",
"base_dir": "data/irish_core_diffmask_v5_mix",
"seed": 20260315,
"train_rows": 7413,
"valid_rows": 932,
"test_rows": 924,
"focus": "GlobalPointer extended-label mix: current chatbot focus plus Irish address/city/county data, new DOB/AGE demographic patch, and multilingual PPSN preservation."
},
"source_patch_sets": [
"data/irish_core_globalpointer_focus_chatbot_v3",
"data/irish_other_pii_v1_mix",
"data/globalpointer_demographic_patch_v1",
"data/irish_core_diffmask_v5_mix"
],
"evaluation_correction_note": "globalpointer_demographic_patch_v2_test uses the corrected Eircode generator; v1 included invalid Eircodes in some synthetic rows.",
"policy_eval_suites": [
"eval/irish_extended_pii_v1.jsonl",
"eval/irish_gov_contact_policy_v1.jsonl",
"eval/irish_gov_chatbot_redteam_v2.jsonl",
"eval/irish_gov_chatbot_gap_holdout_v2.jsonl",
"eval/globalpointer_demographic_patch_v2_test.jsonl"
],
"release_delta": {
"from": "IrishCore-GlobalPointer-ContextPII-135M-v1-rc3",
"type": "decoder_only",
"details": "Fixes contextless sort-code false positives, short/suffixed phone false positives, restores Dublin landline recall, and tightens city/org-prefix handling without changing weights or ONNX graphs."
}
}