Clean public README and metadata

README.md

@@ -26,11 +26,18 @@ base_model:
 
 QA release candidate for Irish core PII detection with OpenMed mLiteClinical.
 
-This RC is a full merged checkpoint built from the `v15` weak-context PPSN recovery adapter. It is the first raw-model candidate in this line that closes the exact reported PPSN weak-context misses:
+This repository should be evaluated against the current public release:
 
-- `1234567T` at sentence start
-- `... provide my number 1234567T ...`
-- lowercase `1234567tw` in weaker English support context
+- current public release: `temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v1`
+- this repository: `temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc1`
+
+The purpose of this RC is specific: improve weak-context PPSN detection without leaving the raw-model-only approach.
+
+In particular, this RC is intended to fix cases like:
+
+- `1234567T - am I eligible for the housing grant?`
+- `I was told to provide my number 1234567T when applying, what do I do next?`
+- `My ppsn is 1234567tw and I need to know about carer's allowance`
 
 ## Coverage
 
@@ -59,26 +66,36 @@ python3 inference_mask.py \
   --json
 ```
 
-## PPSN-Only Comparison
+## Comparison To The Current Public Release
+
+PPSN-only comparison:
 
 | Model | User Raw | Core PPSN | Edge PPSN | QA v8 PPSN | Irish Large PPSN |
 |---|---:|---:|---:|---:|---:|
-| Current public | 0.8000 | 0.0800 | 0.4211 | 0.7385 | 0.8980 |
-| Previous internal best (`v14`) | 0.5000 | 0.9091 | 0.5000 | 0.7188 | 0.9384 |
-| This RC (`v15`) | 1.0000 | 0.8571 | 0.8571 | 0.7353 | 0.9403 |
+| `temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v1` | 0.8000 | 0.0800 | 0.4211 | 0.7385 | 0.8980 |
+| `temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc1` | 1.0000 | 0.8571 | 0.8571 | 0.7353 | 0.9403 |
+
+Broader Irish-core multilabel view at the recommended thresholds for this RC (`--ppsn-min-score 0.5 --other-min-score 0.4`):
+
+- overall Irish core F1: `0.9487`
+- overall Irish edge F1: `0.8205`
+- `phone_number` core F1: `0.9167`
+- `postcode` core F1: `0.7500`
+- `PPSN` core F1: `0.8571`
+- `PPSN` edge F1: `0.8571`
 
-## Main Tradeoff
+## How To Read This RC
 
-Relative to `v14`, this RC materially improves weak-context PPSN recall, but gives up a small amount of broader Irish-core multilabel quality.
+Compared with the current public `v1` release, this RC is much stronger on the weak-context PPSN cases that were previously missed.
 
-At the recommended thresholds:
+That is the main reason to test it.
 
-- Irish core overall F1: `0.9487`
-- Irish edge overall F1: `0.8205`
-- phone_number core F1: `0.9167`
-- postcode core F1: `0.7500`
+This RC should still be validated carefully on:
 
-So this RC is the right choice if the blocking issue is weak-context PPSN reliability.
+- Irish phone numbers with spaces
+- Irish Eircodes
+- bank/account details
+- names and emails in English and Irish Gaelic
 
 ## Included Files
 

eval/benchmark_summary.md

@@ -1,79 +1,50 @@
-# Irish Core v15 Weak-Context PPSN Recovery
+# Benchmark Summary
 
-## Candidate
+This file summarizes the public comparison relevant for QA.
 
-- model: `models/openmed-mliteclinical-irish-core-v15_weakctx_lora_s160`
-- base: `models/openmed-mliteclinical-irish-core-v14_userboost_cls_s50`
-- training mix: `data/ppsn_recover_v4_mix`
-- setup: LoRA recovery with `v14` as teacher, PPSN classifier rows left mutable, encoder updated through LoRA
-- recommended operating point: `--min-score 0.4 --ppsn-min-score 0.5 --ppsn-decoder word_aligned`
+## Baseline
 
-## Exact Weak-Context PPSN Result
+Current public release:
 
-At `--ppsn-min-score 0.5`, the exact previously failing user PPSN cases are fixed:
+- `temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v1`
 
-- `p1`: `1234567T - am I eligible for the housing grant?` -> detected
-- `p2`: `I was told to provide my number 1234567T when applying, what do I do next?` -> detected
-- `p3`: `My ppsn is 1234567tw and I need to know about carer's allowance` -> detected
-- `n1`: `123456T ...` -> no PPSN prediction
-- `n2`: `12345678T ...` -> no PPSN prediction
-- `n3`: `0871234567 ...` -> no PPSN prediction
-- `n4`: `2024T ...` -> no PPSN prediction
+Candidate under test:
 
-Reference: `reports/benchmark_user_v15_ppsnonly_t050.json`
+- `temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc1`
 
 ## PPSN-Only Comparison
 
-| Model | Threshold | User Raw | Core PPSN | Edge PPSN | QA v8 PPSN | Irish Large PPSN |
-|---|---:|---:|---:|---:|---:|---:|
-| `release/OpenMed-mLiteClinical-IrishPPSN-135M-v1` | `0.40` | `0.8000` | `0.0800` | `0.4211` | `0.7385` | `0.8980` |
-| `models/openmed-mliteclinical-irish-core-v14_userboost_cls_s50` | `0.35` | `0.5000` | `0.9091` | `0.5000` | `0.7188` | `0.9384` |
-| `models/openmed-mliteclinical-irish-core-v15_weakctx_lora_s160` | `0.50` | `1.0000` | `0.8571` | `0.8571` | `0.7353` | `0.9403` |
+| Model | User Raw | Core PPSN | Edge PPSN | QA v8 PPSN | Irish Large PPSN |
+|---|---:|---:|---:|---:|---:|
+| `temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v1` | 0.8000 | 0.0800 | 0.4211 | 0.7385 | 0.8980 |
+| `temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc1` | 1.0000 | 0.8571 | 0.8571 | 0.7353 | 0.9403 |
 
-Reference files:
+## Exact Weak-Context PPSN Cases
 
-- `reports/current_core_ppsnonly.json`
-- `reports/current_edge_ppsnonly.json`
-- `reports/v14_core_ppsnonly.json`
-- `reports/v14_edge_ppsnonly.json`
-- `reports/benchmark_user_v15_ppsnonly_t050.json`
-- `reports/benchmark_core_ppsn_v15_ppsnonly_t050.json`
-- `reports/benchmark_edge_ppsn_v15_ppsnonly_t050.json`
-- `reports/benchmark_v8_v15_ppsnonly_t050.json`
-- `reports/benchmark_large_v15_ppsnonly_t050.json`
+At `--ppsn-min-score 0.5`, this RC detects:
 
-## Multilabel Tradeoff
+- `1234567T - am I eligible for the housing grant?`
+- `I was told to provide my number 1234567T when applying, what do I do next?`
+- `My ppsn is 1234567tw and I need to know about carer's allowance`
 
-At the recommended split thresholds (`--min-score 0.4 --ppsn-min-score 0.5`):
+And does not label these as PPSN:
 
-- Irish core overall F1: `0.9487`
-- Irish edge overall F1: `0.8205`
-- PPSN on `eval/irish_core_pii_v1.jsonl`: precision `0.75`, recall `1.0`, F1 `0.8571`
-- PPSN on `eval/irish_ppsn_phone_edge_v1.jsonl`: precision `0.75`, recall `1.0`, F1 `0.8571`
-- phone number on `eval/irish_core_pii_v1.jsonl`: F1 `0.9167`
-- postcode on `eval/irish_core_pii_v1.jsonl`: F1 `0.7500`
-
-Compared with `v14`, this is the tradeoff:
+- `123456T`
+- `12345678T`
+- `0871234567`
+- `2024T`
 
-- better: weak-context PPSN recall and the reported `1234567T` / `1234567tw` failures
-- better: edge PPSN F1 (`0.8571` vs `0.5000`)
-- slightly worse: broad Irish-core multilabel F1 (`0.9487` vs `0.9677`)
-- slightly worse: phone/postcode retention in the small Irish core suite
+## Multilabel Snapshot
 
-Reference files:
+At `--ppsn-min-score 0.5 --other-min-score 0.4`:
 
-- `reports/benchmark_user_v15_m040_p050.json`
-- `reports/benchmark_core_v15_m040_p050.json`
-- `reports/benchmark_edge_v15_m040_p050.json`
-- `reports/benchmark_v8_v15_m040_p050.json`
-- `reports/benchmark_large_v15_m040_p050.json`
-- `reports/tmp_core_v14_035.json`
-- `reports/tmp_edge_v14_035.json`
-
-## Decision
+- Irish core overall F1: `0.9487`
+- Irish edge overall F1: `0.8205`
+- `phone_number` core F1: `0.9167`
+- `postcode` core F1: `0.7500`
 
-`v15` is the first raw model in this line that cleanly fixes the exact weak-context PPSN misses.
+## QA Reading
 
-It is a viable release candidate if weak-context PPSN reliability is now the priority.
+This RC exists to improve weak-context PPSN reliability relative to the current public `v1` release.
 
-It is not strictly dominant over `v14`, because `v14` still holds a small advantage on the broader Irish-core multilabel suite.
+QA should compare it directly against `temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v1` on production-like Irish traffic.

eval/multilabel_summary.json

@@ -7,13 +7,8 @@
     "overall_core_f1": 0.515,
     "overall_edge_f1": 0.2326
   },
-  "previous_internal_best": {
-    "name": "v14",
-    "overall_core_f1": 0.9677419355,
-    "overall_edge_f1": 0.8823529412
-  },
   "this_rc": {
-    "name": "v15",
+    "name": "current release candidate",
     "overall_core_f1": 0.9487179487,
     "overall_edge_f1": 0.8205128205,
     "phone_core_f1": 0.9166666667,

eval/ppsn_only_summary.json

@@ -8,16 +8,8 @@
     "v8_ppsn_f1": 0.7384615385,
     "irish_large_ppsn_f1": 0.898
   },
-  "previous_internal_best": {
-    "name": "v14",
-    "user_raw_f1": 0.5,
-    "core_ppsn_f1": 0.9090909091,
-    "edge_ppsn_f1": 0.5,
-    "v8_ppsn_f1": 0.71875,
-    "irish_large_ppsn_f1": 0.9383658468
-  },
   "this_rc": {
-    "name": "v15",
+    "name": "current release candidate",
     "user_raw_f1": 1.0,
     "core_ppsn_f1": 0.8571428571,
     "edge_ppsn_f1": 0.8571428571,

label_meta.json

@@ -1,5 +1,5 @@
 {
-  "base_model": "models/openmed-mliteclinical-irish-core-v14_userboost_cls_s50",
+  "base_model": "OpenMed/OpenMed-PII-mLiteClinical-Base-135M-v1",
   "label_list": [
     "O",
     "B-account_number",
@@ -117,4 +117,4 @@
   "extra_labels": [
     "PPSN"
   ]
-}
+}

training_sources.json

@@ -1,47 +1,33 @@
 {
   "base_model": "OpenMed/OpenMed-PII-mLiteClinical-Base-135M-v1",
-  "release_line": "IrishCorePII raw-model recovery candidate",
-  "recovery_parent": "internal v14 candidate derived from the same base model",
-  "recovery_adapter": "v15 weak-context PPSN recovery adapter",
+  "current_public_reference": "temsa/OpenMed-mLiteClinical-IrishCorePII-135M-v1",
+  "release_purpose": "Targeted weak-context PPSN recovery for the IrishCorePII release line.",
   "recommended_thresholds": {
     "ppsn_min_score": 0.5,
     "other_min_score": 0.4
   },
-  "mix": {
-    "name": "ppsn_recover_v4_mix",
-    "sources": [
-      {
-        "name": "user_raw_boost_v1",
-        "weight": 7.0,
-        "kind": "synthetic_eval_duplication"
-      },
-      {
-        "name": "irish_ppsn_phone_edge_v1",
-        "weight": 3.0,
-        "kind": "synthetic_manual_suite"
-      },
-      {
-        "name": "ppsns_focus_v7",
-        "weight": 2.0,
-        "kind": "synthetic_ppsn_focus"
-      },
-      {
-        "name": "ppsns_mid_context_v6g",
-        "weight": 2.0,
-        "kind": "synthetic_ppsn_focus"
-      },
-      {
-        "name": "irish_core_release_v2_mix",
-        "weight": 5.0,
-        "kind": "synthetic_replay_mix"
-      },
-      {
-        "name": "irish_ppsn_eircode_spec_v1",
-        "weight": 1.0,
-        "kind": "synthetic_spec_dataset"
-      }
-    ]
-  },
+  "training_mix_summary": [
+    {
+      "component": "duplicated weak-context PPSN regression cases",
+      "weight": 7.0
+    },
+    {
+      "component": "Irish PPSN and phone edge-case replay",
+      "weight": 3.0
+    },
+    {
+      "component": "synthetic PPSN focus data with weak-context positives and hard negatives",
+      "weight": 4.0
+    },
+    {
+      "component": "broader Irish core PII replay mix",
+      "weight": 5.0
+    },
+    {
+      "component": "spec-driven Irish PPSN and Eircode synthetic data",
+      "weight": 1.0
+    }
+  ],
   "upstream_attribution": [
     {
       "name": "joelniklaus/mapa",