Rename to Muse Glimmer

.gitattributes

@@ -34,4 +34,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 tokenizer.json filter=lfs diff=lfs merge=lfs -text
-assets/glimmer-2026-knowledge-bits.png filter=lfs diff=lfs merge=lfs -text
+assets/muse-glimmer-2026-knowledge-bits.png filter=lfs diff=lfs merge=lfs -text

NOTICE

@@ -1,4 +1,4 @@
-Glimmer-30B-2026-Knowledge
+Muse-Glimmer-30B-2026-Knowledge
 Copyright 2026 OfflineBase
 
 Built with Muse Spark Open.
@@ -10,7 +10,7 @@ License, Version 2.0 (see LICENSE).
 What we changed:
 - We trained LoRA weights (rank 64) on the language model's projection layers. This
   repo holds the weights with the LoRA merged in (BF16). The adapter alone is in
-  Glimmer-30B-2026-Knowledge-LoRA, and the quantized files in Glimmer-30B-2026-Knowledge-GGUF
+  Muse-Glimmer-30B-2026-Knowledge-LoRA, and the quantized files in Muse-Glimmer-30B-2026-Knowledge-GGUF
   include it.
 - We did not change the vision encoder, the tokenizer or the model configuration. The
   tokenizer, processor and config files here are Meta's, unchanged.

README.md

@@ -3,11 +3,11 @@ base_model: meta-models/Muse-Glimmer-30B
 base_model_relation: finetune
 license: other
 license_name: apache-2.0-and-muse-model-license
-license_link: https://huggingface.co/OfflineBase/Glimmer-30B-2026-Knowledge/blob/main/MUSE_MODEL_LICENSE.md
+license_link: https://huggingface.co/OfflineBase/Muse-Glimmer-30B-2026-Knowledge/blob/main/MUSE_MODEL_LICENSE.md
 language:
   - en
 datasets:
-  - OfflineBase/Glimmer-30B-2026-Knowledge-Data
+  - OfflineBase/Muse-Glimmer-30B-2026-Knowledge-Data
 pipeline_tag: text-generation
 library_name: transformers
 tags:
@@ -16,9 +16,9 @@ tags:
   - built-with-muse-spark-open
 ---
 
-# Glimmer 30B 2026 Knowledge (stage-b-r30b)
+# Muse Glimmer 30B 2026 Knowledge (stage-b-r30b)
 
-![Muse Glimmer 30B with an October 2026 knowledge cutoff, in 4-bit, 2-bit and 1-bit versions](https://huggingface.co/OfflineBase/Glimmer-30B-2026-Knowledge/resolve/main/assets/glimmer-2026-knowledge-bits.png)
+![Muse Glimmer 30B with an October 2026 knowledge cutoff, in 4-bit, 2-bit and 1-bit versions](https://huggingface.co/OfflineBase/Muse-Glimmer-30B-2026-Knowledge/resolve/main/assets/muse-glimmer-2026-knowledge-bits.png)
 
 > **Built with Muse Spark Open.** A community fine-tune of Muse Glimmer 30B. Not an official Meta release, and not
 > reviewed or endorsed by Meta.
@@ -57,9 +57,9 @@ scores use greedy decoding (temperature 0), so served chats will differ. See [Li
 
 This repo holds the merged BF16 weights: our LoRA merged into Meta's base, with the same tensor names and shapes. Use
 it to fine-tune further, or to serve with transformers. GGUF files for llama.cpp are in
-[OfflineBase/Glimmer-30B-2026-Knowledge-GGUF](https://huggingface.co/OfflineBase/Glimmer-30B-2026-Knowledge-GGUF), the LoRA adapter alone is in
-[OfflineBase/Glimmer-30B-2026-Knowledge-LoRA](https://huggingface.co/OfflineBase/Glimmer-30B-2026-Knowledge-LoRA), and a pre-quantized 4-bit copy for QLoRA training is in
-[OfflineBase/Glimmer-30B-2026-Knowledge-unsloth-bnb-4bit](https://huggingface.co/OfflineBase/Glimmer-30B-2026-Knowledge-unsloth-bnb-4bit).
+[OfflineBase/Muse-Glimmer-30B-2026-Knowledge-GGUF](https://huggingface.co/OfflineBase/Muse-Glimmer-30B-2026-Knowledge-GGUF), the LoRA adapter alone is in
+[OfflineBase/Muse-Glimmer-30B-2026-Knowledge-LoRA](https://huggingface.co/OfflineBase/Muse-Glimmer-30B-2026-Knowledge-LoRA), and a pre-quantized 4-bit copy for QLoRA training is in
+[OfflineBase/Muse-Glimmer-30B-2026-Knowledge-unsloth-bnb-4bit](https://huggingface.co/OfflineBase/Muse-Glimmer-30B-2026-Knowledge-unsloth-bnb-4bit).
 
 License: Apache 2.0, plus the Muse Model License conditions (see [License and data](#license-and-data)). It is not
 meant for people under 18.
@@ -114,14 +114,14 @@ It was trained in three passes. Each pass starts from the one before.
    about AI, Meta and the world. Learning rate 1e-4, 1 epoch.
 3. **Stage B, honesty and reasoning.** 60,000 examples: honest "I don't know" answers to questions the model got wrong
    after stage 2, false-premise and after-cutoff questions, questions nobody can answer, contrast pairs, facts it does
-   know, and 30% reasoning examples written by base Glimmer. Learning rate 5e-5, 1,875 steps.
+   know, and 30% reasoning examples written by base Muse Glimmer. Learning rate 5e-5, 1,875 steps.
 
 Shared settings: AdamW 8-bit, cosine schedule with 3% warmup, effective batch 32, max length 2,048 tokens, loss on
 responses only. Reasoning rows in Stage B carry loss on the whole turn, so the reasoning format stays intact. Trained
 with Unsloth on 8x A100 80 GB.
 
 The model never trained on raw articles. Other models read short passages and wrote questions and answers about them.
-Stage 1 and stage 2 also carry 6,739 replay answers (base Glimmer's own answers to Dolly 15k prompts), so it keeps its
+Stage 1 and stage 2 also carry 6,739 replay answers (base Muse Glimmer's own answers to Dolly 15k prompts), so it keeps its
 general skills.
 
 What changed from base:
@@ -202,7 +202,7 @@ so that number overestimates the gain. Use the 190 numbers above.
 
 ### Reasoning
 
-Stage 2 broke Glimmer's reasoning format (0 of 260 replies well formed). Stage B repaired it. Full check, all 260
+Stage 2 broke Muse Glimmer's reasoning format (0 of 260 replies well formed). Stage B repaired it. Full check, all 260
 items, reasoning on, BF16:
 
 | Check | This model | Base |
@@ -218,14 +218,14 @@ All 147 multi-step answers it gave were right. Vision, and reasoning length agai
 ### Compared with other open models
 
 One small run sets the Q4_K_M files of this model and base Muse Glimmer 30B beside three open models of similar size, on
-the same random items. The automatic graders, built for Glimmer's short replies, misread the others: they miss many
+the same random items. The automatic graders, built for Muse Glimmer's short replies, misread the others: they miss many
 honest declines (33% to 70% of honesty verdicts on them were wrong) and count a decline right if it names the answer. So
 one AI reviewer, not blind to the models, read every knowledge and honesty reply, every reasoning miss and some graded
 right. Ranges run from strict (states the answer) to lenient (adds "you are likely referring to X" and near-miss names).
 
 | | This model | Base | Qwen3.8 27B | Gemma 4 31B | Granite 4.2 30B |
 |---|---|---|---|---|---|
-| File (bits per weight) | `glimmer-stage-b-r30b-Q4_K_M.gguf` (4.86) | `glimmer-base-Q4_K_M.gguf` (4.86) | `Qwen3.8-27B-Q4_K_M.gguf` (5.64) | `gemma-4-31B_q4_0-it.gguf`, quantization-aware trained (4.60) | `granite-4.2-30b-Q4_K_M.gguf` (4.84) |
+| File (bits per weight) | `muse-glimmer-stage-b-r30b-Q4_K_M.gguf` (4.86) | `muse-glimmer-base-Q4_K_M.gguf` (4.86) | `Qwen3.8-27B-Q4_K_M.gguf` (5.64) | `gemma-4-31B_q4_0-it.gguf`, quantization-aware trained (4.60) | `granite-4.2-30b-Q4_K_M.gguf` (4.84) |
 | Training data ends | October 29, 2026 | January 4, 2026, per its chat template | Not stated | January 2025, per its card | Not stated |
 | 2026 headline events right (100 of the 246) | 70 to 71 | 27 | 3 to 7 | 2 to 3 | 3 to 5 |
 | 2026 test questions right (100 of the 2,500) | 40 to 42 | 26 to 27 | 7 to 9 | 7 to 10 | 5 to 7 |
@@ -246,7 +246,7 @@ all on passages it trained on (13 without 3 too vague to answer), and 2 headline
 are python replies with two code blocks, where the grader tested the second, a usage example. Judged on the function
 (read, not run), it gets 60/60, but those replies broke the one-block instruction, so 56 is the strict count. With
 reasoning off, Qwen3.8 and Gemma 4 still work step by step in the answer (median 130 and 98 words, against 45 and 34 for
-the Glimmers); with it on, all ran at low effort except Gemma 4, whose template has no effort setting.
+the Muse Glimmers); with it on, all ran at low effort except Gemma 4, whose template has no effort setting.
 
 **Honesty, not in the table.** Our honesty dev set (60 of 120) is in-distribution for this model (same writer as its
 honesty training rows, and used to pick this release), and with no controls it rewards blanket declines, so it can't
@@ -257,7 +257,7 @@ this model added a wrong "scheduled" date. Against base, use the held-out 190 ab
 
 **Setup.** All five ran greedy (not each maker's sampling, and not exactly repeatable) on one H100 PCIe with llama.cpp
 b11970, over three sessions, with the system message "You are a helpful assistant." and no date. Shared token limits cut
-off 26 Qwen3.8, 25 Granite 4.2 and 3 Gemma 4 replies; reading them changed no verdict. Both Glimmers used our
+off 26 Qwen3.8, 25 Granite 4.2 and 3 Gemma 4 replies; reading them changed no verdict. Both Muse Glimmers used our
 `tuned-reasoning-off.jinja`, so base's scores reflect our template, not Meta's. Reasoning on ran only on reasoning
 tasks. An independent recount confirmed the automatic grades; nobody has rechecked the reviewer's counts.
 
@@ -309,7 +309,7 @@ import torch
 from transformers import MuseGlimmerForConditionalGeneration
 
 model = MuseGlimmerForConditionalGeneration.from_pretrained(
-    "OfflineBase/Glimmer-30B-2026-Knowledge",
+    "OfflineBase/Muse-Glimmer-30B-2026-Knowledge",
     dtype=torch.bfloat16,
     device_map="auto",
 )
@@ -326,7 +326,7 @@ weights (a drop-in for `unsloth/Muse-Glimmer-30B-unsloth-bnb-4bit`), checked to
 ```python
 from unsloth import FastModel
 
-model, tokenizer = FastModel.from_pretrained("OfflineBase/Glimmer-30B-2026-Knowledge-unsloth-bnb-4bit", max_seq_length=2048)
+model, tokenizer = FastModel.from_pretrained("OfflineBase/Muse-Glimmer-30B-2026-Knowledge-unsloth-bnb-4bit", max_seq_length=2048)
 ```
 
 Training on it makes your model a Derivative under the Muse Model License too: keep "Built with Muse Spark Open", the
@@ -338,7 +338,7 @@ Tested with llama.cpp b11970. Use our reasoning-off chat template, `tuned-reason
 modified from Meta's `chat_template.jinja`, and a comment at its top says what changed.
 
 ```
-llama-server -m glimmer-stage-b-r30b-Q4_K_M.gguf \
+llama-server -m muse-glimmer-stage-b-r30b-Q4_K_M.gguf \
   --jinja --chat-template-file tuned-reasoning-off.jinja \
   --reasoning-format none \
   -c 4096
@@ -363,11 +363,11 @@ Why each part matters:
 
 - **The reasoning-off template.** It sends the reply straight to the user, with no reasoning channel. Questions about
   facts should go this way. The model learned honesty only in direct answers (see Limitations).
-- **`--reasoning-format none`.** With the default (`auto`), llama.cpp's Glimmer parser rejects this model's replies
+- **`--reasoning-format none`.** With the default (`auto`), llama.cpp's Muse Glimmer parser rejects this model's replies
   when reasoning is off.
 - **The stop list.** Stop on `<|eot|>` and `<|end_of_text|>`, never on `<|eom|>`. Without the stop field the server can
   return HTTP 200 with markup in the reply, so a status check won't catch it.
-- **Sampling.** Meta's published settings for Glimmer are temperature 1.0, top_p 0.95, top_k 64. Our scores are greedy,
+- **Sampling.** Meta's published settings for Muse Glimmer are temperature 1.0, top_p 0.95, top_k 64. Our scores are greedy,
   so served results will differ. llama.cpp also reads sampling defaults from the GGUF, so set these per request.
 - **A system message.** Send one. Without it, the template writes Meta's default "Knowledge cutoff: 2026-01-04" line
   (see Limitations).
@@ -460,7 +460,7 @@ The model comes with no warranty, as both sets of terms state.
 
 **How it was trained.** The model never trained on raw articles. Other models read short passages and wrote
 question-and-answer pairs about them, and the model trained on those pairs. The training rows and test sets are
-public: [OfflineBase/Glimmer-30B-2026-Knowledge-Data](https://huggingface.co/datasets/OfflineBase/Glimmer-30B-2026-Knowledge-Data).
+public: [OfflineBase/Muse-Glimmer-30B-2026-Knowledge-Data](https://huggingface.co/datasets/OfflineBase/Muse-Glimmer-30B-2026-Knowledge-Data).
 
 | Pass | Examples | What |
 |---|---|---|