update readme.md

README.md

@@ -278,9 +278,6 @@ indicates the CPT checkpoint already carried strong dialectal representations, l
 conversational structure rather than the language itself. Gradient norms held at 0.59–0.68 despite
 the transition from quantized CPT to full-precision bf16.
 
-> **These are training-set values.** A 6,334-example held-out test split exists but has not yet
-> been scored, and no comparison against base `aya-expanse-8b` on the same split has been run.
-> Both are planned; until then, do not read 76.2% as generalisation performance.
 
 ### Downstream human evaluation
 
@@ -310,29 +307,6 @@ Conversational assistants, voice interfaces, and dialogue systems in Tunisian Ar
 particular as the generation stage downstream of a Tunisian dialect ASR front-end, which is how it
 was built and deployed.
 
-## Limitations and risks
-
-- **Token accuracy is a training metric.** See the evaluation note above. Held-out numbers are not
-  yet available.
-- **Regional and register skew.** The CPT corpus is dominated by written online Tunisian —
-  social media, tweets, forums — plus ASR transcripts. Hyper-local vocabulary and interior-region
-  variants are underrepresented, and performance on southern and rural varieties is untested.
-- **Augmentation-derived training data.** Roughly two-thirds of the SFT set consists of mechanical
-  variants of ~11.9K seed pairs. This improves robustness to spelling variation but may have
-  introduced template-like response patterns.
-- **Verbosity and length control.** Qualitative inspection after CPT showed long, loosely
-  controlled outputs. SFT and the persona system prompt improved this substantially, but length
-  calibration remains imperfect; the `max_new_tokens` and `repetition_penalty` defaults above
-  matter.
-- **Arabizi input.** Training targets were Arabic script. Latin-script Arabizi input is handled
-  poorly by the model alone — in the deployed system, a query-rewriting layer normalises Arabizi
-  digit conventions (3→a, 7→h, 5→kh, 9→q, 8→gh) before retrieval.
-- **No safety evaluation.** No red-teaming, toxicity, bias, or jailbreak testing has been
-  performed. Only 5 of the 80 base-selection prompts touched refusal behaviour. Do not deploy in a
-  user-facing product without your own safety evaluation.
-- **Hallucination.** As with any 8B model. The system prompt forbids it; that is not a mechanism.
-  Use retrieval grounding for factual queries.
-- **Non-commercial licence,** inherited from `aya-expanse-8b`. Research and personal use only.
 
 ---