Repository navigation
Fix quick reference: default recipe is Qwen3-0.6B full fine-tune, not Qwen3.5-2B + LoRA - #179
Merged
Merged
Conversation
… Qwen3.5-2B + LoRA
alexkroman
marked this pull request as ready for review
November 17, 2026 14:28
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this fixes
docs/course/4-quick-reference.mddescribes the current default recipe (configs/config.yaml+configs/training/production.yaml) as Granite + Qwen3.5-2B, frozen, with LoRA (~80M trained), and lists thefreeze_language_model,freeze_text_embed_tokens, andmodel.use_loradefaults astrue. The shipped config says otherwise:text_model_id: Qwen/Qwen3-0.6B,use_lora: false,freeze_language_model: false,freeze_text_embed_tokens: false— the published model's recipe, ~604M trained parameters (596M decoder + 8.4M projector, per the model card).This corrects the training comment, the Recipes table row, and the three Freeze Flags defaults. Nothing else changes. The course chapters still teach the Qwen3.5-2B frozen recipe as the worked example; whether they should move to the 0.6B recipe is a separate editorial call, left out of this PR on purpose.
The 2B vs 4B question, settled
This audit started from the long-standing flag (Oct 15) that some documents said Qwen3.5-4B while the README and model card said the published model used Qwen3.5-2B. Checking the released weights directly:
granite-qwen-frozen-4ofmazesmazes/tiny-audio, ships aconfig.jsonwhosetext_configisqwen3_5_textwithhidden_size2048 and 24 layers, andtext_model_id: Qwen/Qwen3.5-2B. Qwen3.5-4B is 2560 hidden x 32 layers. The released decoder is 2B.rjid9zj6, 2026-10-16), recorded inconfigs/training/production.yaml's comments as "kept for a future decoder swap." It was never released.main, 2026-10-23) is a different model — Qwen3-0.6B, verified from its releasedconfig.json(hidden 1024, 28 layers).