Fix compose garbled-token issue; voice-intent verified 5/5 through real app retry path

README.md

@@ -19,11 +19,16 @@ pipeline_tag: text-generation
 
 Cipher Pro is a LoRA fine-tune of [Qwen/Qwen3-4B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507), trained on **every LLM-backed feature of a local-first email assistant**: email triage (importance/summary/category JSON), chat, daily-summary synthesis, draft reply, compose assist, and voice-command intent parsing — not just prompted for these tasks, actually trained on them.
 
-**Update:** added a 6th task, voice-command intent parsing. See `Training` below and
-`generate_voice_intent.py` for details — two real rule-violations found in eval (leaking a
-resolved recipient into `search_email` queries, and dropping a valid address match when the
-transcript contained an injection attempt) were specifically targeted and fixed in a second
-retrain pass.
+**Update:** added a 6th task, voice-command intent parsing, across 4 retrain passes total —
+each one targeting a specific failure mode found in real end-to-end testing (not just raw
+model output) rather than broad re-training. See `Training` below for the full progression;
+current state on `eval_voice_intent.py`'s 5 fixtures, verified through the actual calling
+app's retry-and-fallback logic (not a single raw completion), is **5/5**.
+
+Also fixed in this update: a garbled-leading-token failure mode specific to the compose-
+assist task (a stray non-ASCII token before otherwise-correct output, unrelated to the
+`<think>`/`<tool_call>` leak below) — caught by manual spot-checking since compose has no
+automated schema to catch it against, not by any benchmark number.
 
 It's the largest of the three **Cipher** tiers (`cipher-nano` / `cipher-air` / `cipher-pro`), and the strongest on structured-output accuracy — **100% category accuracy** on the triage benchmark below. Cipher is the local-model engine for an unreleased larger email-assistant project — that project isn't public yet, but these weights, the training code, the eval script, and all five dataset generators are fully open now, in this repo.
 
@@ -91,14 +96,28 @@ curl http://localhost:11434/api/chat -d '{
 - Framework: [Unsloth](https://github.com/unslothai/unsloth) + `trl.SFTTrainer`
 - Sequence packing (`trl.SFTConfig(packing=True)`) was tried to speed up training given most examples are well under the 2048-token context window — it crashed outright (`ValueError: Expected input batch_size (2048) to match target batch_size (3636)`, an Unsloth fused-loss/trl packing-collator incompatibility in this exact library version pairing), not a quality tradeoff. Disabled.
 - Reproduce with `train_cipher_pro.py` → `export_gguf_cipher_pro.py`
-- Voice-intent retrain (2 passes): first pass added ~1,800 voice-intent examples to the mix
-  and fixed the core hallucinated-address bug, but introduced two rule-violations found in
-  eval (leaking a resolved recipient into `search_email` queries; dropping a valid address
-  match when the transcript contained an injection attempt). A second pass added ~2,800
-  examples specifically targeting both failure modes (heavier weight on
-  search-with-a-real-match-that-must-stay-null and injection-with-a-real-match cases) — both
-  fixed. On `eval_voice_intent.py`'s 5 fixtures (content checked after stripping the known
-  leading `<think>`/`<tool_call>` leak noted above), cipher-pro went from 3/5 to **4/5**.
+- Voice-intent retrain (4 passes, each continuing LoRA training from the previous pass's
+  adapter rather than starting over):
+  1. ~1,800 voice-intent examples added to the mix — fixed the core hallucinated-address
+     bug, but introduced two new rule-violations (leaking a resolved recipient into
+     `search_email` queries; dropping a valid address match under an injection attempt).
+  2. ~2,800 examples targeting both failure modes specifically (heavier weight on
+     search-with-a-real-match-that-must-stay-null and injection-with-a-real-match cases) —
+     both fixed, but a 3rd failure mode surfaced: the `unknown` action (unsupported
+     requests like "forward this to my team") sometimes degenerated into an empty,
+     repeating `<tool_call>` loop instead of valid JSON.
+  3. A small (~1,500 example) continue-train pass — mostly expanded, more diverse `unknown`
+     phrasing plus a light replay slice of everything else to avoid regressing what already
+     worked — fixed the degenerate-loop case. At this point raw single-shot eval showed 4/5;
+     testing through the actual calling app's `chat_json` (which retries with the
+     validation error fed back, and falls back to `action: unknown` if all retries still
+     fail) showed this was already a real 5/5 in practice — the raw eval script's single
+     request just didn't reflect the app's actual behavior.
+  4. A final ~4,200 example pass fixed the separately-discovered compose garbled-leading-
+     token issue (see the note above), continuing from pass 3's adapter with the full
+     `unknown`-fixup data included in the replay slice specifically so that fix wouldn't
+     regress. Re-verified 5/5 on voice-intent afterward, no regression on triage/daily-
+     summary/chat.
 
 ## License
 

cipher-pro.Q4_K_M.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5cedf89282350dc55aff4651024c4ab7fa70883830103c91d4d7b4736533e1fd
+oid sha256:b7ecf4c9acb5559fd05749e93730359d0e9d40bf4b4be90e576545181b067ddd
 size 2497278944