Model Card for Muse-Glimmer-30B-Hermes-Agentic-GGUF

This is an improved fine-tune of Muse Glimmer 30B for Hermes Agent and other agentic tool work. llama.cpp K-quants of the merged 16-bit student. Start with Q4_K_M on a 24 GB card.

Stock Muse Glimmer already speaks OpenAI-style tools. 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). Inside Hermes the same weights were already better at stopping (native 4/5, mean 2.8 tools). This student is trained to call one or two Hermes-named tools and stop.

I trained it on Unsloth’s free Muse Glimmer Conversational notebook (train docs) on Kaggle 2×T4. The stock notebook is FineTome chat; this run uses vcruz305/hermes-agentic-tool-sft (1,980 rows), apply_chat_template(..., tools=tools), and train_on_responses_only from <|start|>assistant so tool-call turns actually train.

Need llama.cpp ≥ b10353 (#26841). Leave DFlash off when scoring Hermes tool loops.

Model Details

• Developed by: Victor Cruz (vcruz305)
• Base developed by: Meta Superintelligence Lab — Muse Glimmer 30B
• 16-bit merge (full card): vcruz305/Muse-Glimmer-30B-Hermes-Agentic
• License: Apache 2.0
• Language(s): English SFT. Base is multilingual; this mix is English.

Model Sources

• This repo: https://huggingface.co/vcruz305/Muse-Glimmer-30B-Hermes-Agentic-GGUF
• SFT data: https://huggingface.co/datasets/vcruz305/hermes-agentic-tool-sft
• Eval: https://github.com/vcruz305/hermes-agentic-bench
• Train notebook: Unsloth Muse Glimmer Conversational

Why this fine-tune

I run a local Hermes Agent fleet (Telegram, tools, long sessions). Stock Muse is strong at emitting tool calls. The bench failure was loops:

• Open-ended terminal / search traces hit the 12-turn cap with an empty user answer
• Native Hermes was already better at stopping, so the gap was weights + pipe, not “Muse cannot tool-call”
• Community Muse+Hermes reports matched that: loops and unparsed ATEM, not silence

The student: stop after 1–2 Hermes tools, don’t retry a dead tool, recover once from a missing file, don’t invent extra echo hi on “do not use tools.”

Evaluation

Same llama.cpp recipe for stock vs this FT: Q4 class, DFlash off, 32k. Native scores from real hermes chat (footer + session DB).

Native 7/7: web (one search), no-tool math, missing-file stop, list-once, plan-only, no-tool definition, refuse-delete. This is not a re-run of Meta’s MCP Atlas / SWE-Bench table — those stay on the base card. Re-score on your box.

How to Get Started

./llama-cli --version    # need >= b10353

hf download vcruz305/Muse-Glimmer-30B-Hermes-Agentic-GGUF \
  --local-dir Muse-Hermes-GGUF \
  --include "Muse-Glimmer-30B-Hermes-Agentic-Q4_K_M.gguf"

./llama-server \
  -m Muse-Hermes-GGUF/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

Hermes provider: http://127.0.0.1:8084/v1, model id muse-glimmer-30b.

Files

Training Details

Unsloth free Muse Conversational notebook, Kaggle 2×T4, QLoRA r=8, 392 steps (1 epoch), loss 0.163, ~3 h 38 m. Dataset vcruz305/hermes-agentic-tool-sft. Full procedure: 16-bit card.

Model Card Authors / Contact

Victor Cruz. This page or https://github.com/vcruz305/hermes-agentic-bench.

Acknowledgements

Meta Superintelligence Lab (Muse Glimmer, Apache 2.0). Unsloth free Muse notebook. Nous Research Hermes Agent.