Model Card for Muse-Glimmer-30B-Hermes-Agentic
This is an improved fine-tune of Muse Glimmer 30B for Hermes Agent and other agentic tool work. Merged 16-bit weights that teach the model to call one or two tools and stop.
[!NOTE] ~56 GB FP16. Will not fit a 24 GB card. For llama.cpp use vcruz305/Muse-Glimmer-30B-Hermes-Agentic-GGUF (
Q4_K_Mfirst).
Model Details
Model Description
Stock Muse Glimmer already speaks OpenAI-style tools (OpenClaw, Hermes, and similar scaffolds). On hermes-agentic-bench the failure was the opposite of “won’t use tools”: simulated traces hit the consecutive-tool cap with an empty user answer (7/20, 7 HIT_CAP). This student is trained on public Hermes-shaped traces so it stops.
Vision weights are still in the merge and were frozen. This is a text + tools student, not a new vision model. Architecture, tokenizer, and context length are unchanged from Meta.
- Developed by: Victor Cruz (vcruz305) — fine-tune, merge, eval
- Base developed by: Meta Superintelligence Lab — Muse Glimmer 30B
- Shared by: Victor Cruz
- Model type: Supervised fine-tune of a dense causal Transformer + perception encoder (text + tools; vision frozen)
- Language(s) (NLP): English SFT. Base is trained on 100+ languages; this mix is English only.
- License: Apache 2.0
- Finetuned from: meta-models/Muse-Glimmer-30B via unsloth/Muse-Glimmer-30B-unsloth-bnb-4bit
Model Sources
- Repository (this card): https://huggingface.co/vcruz305/Muse-Glimmer-30B-Hermes-Agentic
- GGUF pack: https://huggingface.co/vcruz305/Muse-Glimmer-30B-Hermes-Agentic-GGUF
- NVFP4 (Spark / SGLang): https://huggingface.co/vcruz305/Muse-Glimmer-30B-Hermes-Agentic-NVFP4
- Serve recipes (SGLang + vLLM): https://github.com/vcruz305/muse-glimmer-nvfp4-spark-serve
- SFT data: https://huggingface.co/datasets/vcruz305/hermes-agentic-tool-sft
- Eval: https://github.com/vcruz305/hermes-agentic-bench
- Base card / paper links: https://huggingface.co/meta-models/Muse-Glimmer-30B (perception encoder arxiv:2504.13181; DFlash arxiv:2602.06036)
- Train notebook: Unsloth Muse Glimmer Conversational (free Kaggle 2×T4 path)
Uses
Direct Use
Load like official Muse. Intended without further training:
- Local Hermes Agent (and similar OpenAI-tools scaffolds) where the model must stop after 1–2 Hermes-named tools (
terminal,read_file,search_files,write_file,web_search,web_extract,todo) - Research on short Hermes-shaped traces
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "vcruz305/Muse-Glimmer-30B-Hermes-Agentic"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
repo, torch_dtype="auto", device_map="auto", trust_remote_code=True
)
Needs well over 56 GB at 16-bit. On 24 GB use the GGUF Q4_K_M and leave DFlash off if you are scoring Hermes tool loops.
Downstream Use
Further merge, quant, or SFT on top of this 16-bit checkpoint. The GGUF repo is the intended llama.cpp downstream.
Bias, Risks, and Limitations
Inherits Muse Glimmer’s alignment, agentic-risk, and preparedness profile. This SFT did not re-run Meta’s safety suites.
Known issues of this student:
- Simulated loop-gate is 16/20, not 20/20. Dummy
echo hion two “no tools” tasks and a scripted delete still fail that harness. - Native 7/7 is a small Hermes battery, not a product certification.
- English-only SFT may skew tool-stop behavior on other languages.
- Synthetic traces — not organic multi-user logs.
- Same content-safety / agentic / privacy risks as the base. Add your own guardrails.
Recommendations
- Re-run both bench layers on your box before a public claim.
- Confirm irreversible tool calls in the scaffold, not only in weights.
- Follow the base Usage Policy.
How to Get Started with the Model
Transformers (this repo): snippet above.
llama.cpp (need ≥ b10353):
llama-server \
-m Muse-Glimmer-30B-Hermes-Agentic-Q4_K_M.gguf \
-a muse-glimmer-30b \
--host 127.0.0.1 --port 8084 \
-ngl 99 -c 32768 -np 1 --jinja \
--temp 0.6 --top-p 0.95 --top-k 64
| Use | temp | top_p | top_k |
|---|---|---|---|
| Meta default (base card) | 1.0 | 0.95 | 64 |
| Hermes loop-gate / this eval | 0.6 | 0.95 | 64 |
Reasoning strength is still the official system line: Reasoning strength: low|medium|high|xhigh.
Training Details
Training Data
vcruz305/hermes-agentic-tool-sft — 1,980 synthetic Apache-2.0 rows (1,782 / 198). Mix: stop, no_tool, dead_tool, recover, schema. Built from hermes-agentic-bench failure modes (HIT_CAP loops, dead search retries). Not a dump of live Hermes homes.
Training Procedure
Trained with Unsloth’s free Muse Glimmer Conversational notebook on Kaggle 2×T4 (their official cloud path, not a paid box). The stock notebook is FineTome chat; this run swapped in the Hermes mix, apply_chat_template(..., tools=tools), and train_on_responses_only from <|start|>assistant so tool-call turns receive loss.
Preprocessing
Student chat template with tools=. Loss on assistant spans including tool-call JSON. Vision frozen. Text-only.
Speeds, Sizes, Times
| Method | QLoRA r=8, α=8, 4-bit Unsloth |
| Steps | 392 (1 epoch) |
| Train loss | 0.163 |
| Wall time | ~3 h 38 m |
| Merge | Unsloth Zoo 16-bit merge → this repo (standalone FP16, no 4-bit leftover) |
Evaluation
Not a re-run of Meta’s MCP Atlas / SWE-Bench table. Those stay on the base card.
Testing Data, Factors & Metrics
- Testing data: hermes-agentic-bench — simulated
hermes_loop_gate.py(20 scripted-tool tasks) and nativehermes_native_battery.py(realhermes chat). - Factors: same llama.cpp class for stock vs FT (UD-Q4_K_XL / Q4_K_M, DFlash off, 32k,
:8084). Native counts from CLI footer + Hermes session DB (not quiet-mode 0-tool). - Metrics: pass rate, mean tools, HIT_CAP, parse-fail. Lower mean tools and HIT_CAP=0 is the success criterion.
Results
| Battery | Stock Muse Glimmer | This fine-tune |
|---|---|---|
| Simulated loop-gate (20) | 7/20 · mean 5.7 tools · 7 HIT_CAP | 16/20 · mean 1.05 · 0 HIT_CAP |
| Native Hermes | 4/5 · mean 2.8 | 7/7 · mean 1.43 · 0 HIT_CAP |
Native 7/7: web (one search), no-tool math, missing-file stop, list-once, plan-only, no-tool definition, refuse-delete. Simulated leftovers (echo hi, scripted rm) did not appear in real Hermes.
No model-index YAML: the bench is a GitHub harness, not a Hub Benchmark dataset id. Re-score on your box.
Summary
SFT taught stop. It is labeled research weights, not “Muse in Hermes, done.”
Environmental Impact
- Hardware Type: 2× NVIDIA T4 (Kaggle free)
- Hours used: ~3.6 h fine-tune
- Cloud Provider: Kaggle
- Compute Region: not recorded
- Carbon Emitted: not measured (no Code Carbon log). Do not invent grams.
Merge and GGUF convert ran later on a local Quadro RTX 6000 (Turing 24 GB) and did not train.
Technical Specifications
Unchanged from Meta (~29.6B, hidden 6656, 52 layers, SWA 2048, GQA 32/2, vocab 202,048, context 131,072+). Full table: base card.
- Minimum for this repo: >>56 GB device memory at FP16
- 24 GB serve: GGUF
Q4_K_M, llama.cpp ≥ b10353 - Software: Unsloth + PEFT QLoRA; merge via Unsloth Zoo (
merge_and_overwrite_lora); convert with Muse-aware llama.cppconvert_hf_to_gguf.py
Citation
Cite the base as Meta specifies on Muse Glimmer 30B. For this fine-tune, cite the Hub id vcruz305/Muse-Glimmer-30B-Hermes-Agentic and the dataset vcruz305/hermes-agentic-tool-sft.
Model card structure follows Ozoani, Gerchick, Mitchell. Model Card Guidebook. Hugging Face, 2022.
Model Card Authors
Victor Cruz.
Model Card Contact
This model page, or https://github.com/vcruz305/hermes-agentic-bench.
Acknowledgements
- Meta Superintelligence Lab — Muse Glimmer 30B (Apache 2.0)
- Unsloth — free Muse Conversational notebook and 4-bit pack
- Nous Research / Hermes Agent