Use OpenSML-152M branding throughout model card

README.md

@@ -22,7 +22,7 @@ An independently trained, English-first small language model built with Apple's
 > after checkpoint selection and release review.
 
 **OpenSML-152M is the project-facing name. The implemented model has exactly
-150,439,168 parameters (150.44M).** Its local research identifier is SML-MLX V2.
+150,439,168 parameters (150.44M).**
 This repository describes one model, not a family of released models.
 
 ## At a Glance
@@ -39,7 +39,7 @@ This repository describes one model, not a family of released models.
 | Cluster | One M3 Ultra and three M4 Max Mac Studios |
 | Communication | Four-node Thunderbolt RDMA ring using JACCL |
 | Planned exposure | 15B total tokens, including the completed pilot |
-| Current status | Continued pretraining; no V2 SFT results or final benchmarks |
+| Current status | Continued pretraining; no SFT results or final benchmarks yet |
 
 ## Architecture
 
@@ -57,7 +57,7 @@ not a claim of a new attention architecture.
 
 The model uses a prose-first mixture streamed from Hugging Face. This replaces
 the earlier OpenWebText description on this page; that description does not apply
-to V2.
+to OpenSML-152M.
 
 | Source | Training token share | Selection |
 | --- | ---: | --- |
@@ -187,8 +187,8 @@ Early generations show repetition, factual errors, and incoherence. Improving
 next-token loss does not establish reliable answers or instruction following.
 
 Independent downstream benchmarks, an untouched final evaluation, full generation
-reviews, and post-training results are **not yet reported for V2**. Results from
-earlier SML models and their SFT branches are not results for this model.
+reviews, and post-training results are **not yet reported for OpenSML-152M**.
+Results from earlier experiments are not results for this model.
 Losses are not directly comparable to other models with different tokenizers or
 validation corpora, even when vocabulary sizes are similar.
 
@@ -199,7 +199,7 @@ experimentation. It is not currently a validated assistant. It may generate
 incorrect, repetitive, biased, offensive, or unsafe text. Do not rely on it for
 high-stakes decisions. Broader capability and safety assessments remain pending.
 
-Architecture, tokenizer, data, and context differ from earlier SML experiments;
+Architecture, tokenizer, data, and context differ from earlier experiments;
 this run alone cannot attribute any improvement to a single change.
 
 ## Repository Contents and Release Plan