release: flowcast-v5-lite v0.2.1 — loudink_surgical + paraphrase_posttrain, 83.3% paraphrase shipped

README.md

@@ -8,7 +8,7 @@ tags:
 library_name: mlx-lm
 ---
 
-# flowcast-v5-lite v0.2.0
+# flowcast-v5-lite v0.2.1
 
 Sub-750MB hybrid LoudInk stack: **99.5% user-demand shipped** (126 computer + 74 writing).
 
@@ -17,7 +17,7 @@ Predecessor: [nsalerni/flowcast-v4-lite](https://huggingface.co/nsalerni/flowcas
 ## Hybrid production stack
 
 - **Writer longform:** `loudink_surgical` — 100% user-demand writing (shipped)
-- **Writer paraphrase:** `onpolicy_rescue` — hybrid routing for held-out paraphrase suite
+- **Writer paraphrase:** `paraphrase_posttrain` — 83.3% shipped / 66.7% neural on held-out paraphrase (12 cases)
 - **IR:** FunctionGemma IR v5 + transcript-first repairs (github-in-safari, etc.)
 - Learned transcript router + polish registry + transcript fast paths
 
@@ -28,6 +28,7 @@ Predecessor: [nsalerni/flowcast-v4-lite](https://huggingface.co/nsalerni/flowcas
 | Combined | 65.5% | **99.5%** |
 | Writing | 9.5% | **100%** |
 | Computer | 98.4% | **99.2%** |
+| Paraphrase (held-out) | — | **83.3%** shipped |
 | Download | 809 MB | **750 MB** |
 
 See `inference_config.json` and `manifest.json` for integration fields.

inference_config.json

@@ -29,8 +29,9 @@
     "ir_fallback_p50_ms_max": 250,
     "neural_writing_accuracy_min": 0.15,
     "neural_computer_accuracy_min": 0.22,
-    "paraphrase_shipped_min": 0.48,
+    "paraphrase_shipped_min": 0.55,
+    "paraphrase_neural_min": 0.3,
     "polish_lift_max": 0.88
   },
-  "version": "0.2.0"
+  "version": "0.2.1"
 }

manifest.json

@@ -1,13 +1,13 @@
 {
   "product": "v5-lite",
-  "version": "0.2.0",
+  "version": "0.2.1",
   "download_mb": 750,
   "runner_kind": "v5_lite",
   "writer_model_config": "configs/model_v5_lfm25_1_2b.yaml",
   "writer_adapter_path": "artifacts/sft_v5_lfm25_v3lite_writer/adapters",
   "writer_checkpoint": "promoted_core_100.safetensors",
   "writer_longform_adapter_path": "artifacts/sft_v5_lite_loudink_surgical/adapters",
-  "writer_paraphrase_adapter_path": "artifacts/sft_v5_lite_onpolicy_rescue/adapters",
+  "writer_paraphrase_adapter_path": "artifacts/sft_v5_lite_paraphrase_posttrain/adapters",
   "hybrid_paraphrase_adapter": true,
   "writer_use_prompt_kv_cache": true,
   "ir_model_config": "configs/model_v4_functiongemma.yaml",
@@ -38,8 +38,9 @@
     "ir_fallback_p50_ms_max": 250,
     "neural_writing_accuracy_min": 0.15,
     "neural_computer_accuracy_min": 0.22,
-    "paraphrase_shipped_min": 0.48,
+    "paraphrase_shipped_min": 0.55,
+    "paraphrase_neural_min": 0.30,
     "polish_lift_max": 0.88
   },
-  "notes": "v5-lite hybrid production: loudink_surgical + onpolicy_rescue paraphrase routing, IR v5, learned router, polish registry."
+  "notes": "v5-lite hybrid production: loudink_surgical + paraphrase_posttrain (83.3% shipped / 66.7% neural paraphrase), IR v5, learned router, polish registry."
 }

writer_paraphrase/adapters.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d426d353fe3740fe564caaf19645055ff10abdd383bdfeb0c2b14cd5fe5fbc23
+oid sha256:8f08dbe5d5aefcfbb36a358750a00941a0e05c3a49a1d666844f694bf9bfdae1
 size 44453426