This is part of a bard model I created with a first draft of an epic fantasy adventure novel I wrote called For The Small Shield, published in GitHub: https://github.com/wordsum/For_The_Small_Shield

The size of the bard model is the smallest I was able to use to generate meaningful output, given the limited amount ot fine-tuing data I had to train the bard LLM, only 2,000 QA pairs: https://huggingface.co/datasets/wordsum/for-the-small-shield-instruct

The bard LLM is not terrible, and the stories of the bard are more historically accurate for the fantasy world when the bard is given the For The Small Shield GraphRAG of Chapters: https://huggingface.co/datasets/wordsum/for-the-small-shield-chapters

The bard is a happy mess of a song that I will revisit in a few months, with all the lessons learned from creating this bard LLM. And I will likely add bonus QA Pairs for the history of Gods and Realms from Of The Small Shield: https://www.instagram.com/ofthesmallshield

Like the original For The Small Shield, Of The Small Shield are stories in dioramas created by me and words written by me, Kalab J. Oster, unlike the words that follow.

Llama-3-for-the-small-shield (v5)

Built with Meta Llama 3

GGUF build of Llama-3-for-the-small-shield (ftss) v5 — a character LoRA fine-tune that role-plays Carlos, the Barded Dwarf, the greatest bard of the fantasy universe Rise&Set. Packaged for local inference with llama.cpp, Ollama, LM Studio, and compatible runtimes.

Model summary

Model name: Llama-3-for-the-small-shield-v5
Developed by: kalaboster (published under the wordsum organization)
Character: Carlos, the Barded Dwarf (universe: Rise&Set)
Model type: decoder-only causal LM (Llama-3 architecture, 8B)
Base model: NousResearch/Hermes-2-Pro-Llama-3-8B
Fine-tuning: LoRA (r=32, α=64), merged into the base, then quantized
Format: GGUF, Q4_K_M (~4.6 GB)
Context length: 8192
Prompt format: ChatML (<|im_start|> / <|im_end|>)
Language: English
Training data: wordsum/for-the-small-shield-instruct (2,001 records)
License: Meta Llama 3 Community License (inherited from the base)

Files

This repo ships the model in two formats:

File(s)	Format	Size	Use
`model-0000{1..4}-of-00004.safetensors` + `config.json` + tokenizer	Safetensors (Transformers)	~15 GB	Load with `from_pretrained`, fine-tune, re-quantize
`Llama-3-for-the-small-shield-v5.Q4_K_M.gguf`	GGUF (Q4_K_M)	~4.6 GB	Local inference: llama.cpp / Ollama / LM Studio

Licence artefacts: LICENSE, NOTICE, USE_POLICY.md.

A standalone LoRA adapter (~321 MB) is published separately at wordsum/for-the-small-shield-lora.

How to run

Transformers (safetensors)

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "wordsum/for-the-small-shield", torch_dtype="bfloat16", device_map="auto"
)
tok = AutoTokenizer.from_pretrained("wordsum/for-the-small-shield")

Ollama

A ready-made Modelfile (with Carlos's system prompt and sampling settings) is included:

bash

ollama create Llama-3-for-the-small-shield -f Modelfile
ollama run Llama-3-for-the-small-shield

llama.cpp

bash

huggingface-cli download wordsum/for-the-small-shield \
  Llama-3-for-the-small-shield-v5.Q4_K_M.gguf --local-dir .

./llama-cli -m Llama-3-for-the-small-shield-v5.Q4_K_M.gguf \
  -c 8192 --temp 0.2 --chat-template-file chat_template.jinja \
  -p "Carlos, tell me the tale of the Small Shield."

LM Studio / Jan

Search wordsum/for-the-small-shield, or load the downloaded .gguf.

Prompt template (ChatML)

<|im_start|>system
{system}<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

Recommended system prompt (as trained):

You are Carlos, the Barded Dwarf... the greatest Bard that exists in the fantasy universe of Rise&Set. (full text in the included Modelfile)

Suggested sampling: temperature 0.2, top_p 0.92, top_k 50, repeat_penalty 1.1.

Training

Base: NousResearch/Hermes-2-Pro-Llama-3-8B
Method: LoRA (r=32, α=64), ~375 steps, then adapter merged into the base and quantized to Q4_K_M GGUF.
Data: for-the-small-shield-instruct — 2,001 {system, input, output} records covering Carlos's voice and the Rise&Set world (characters, places, items, lore).

Data provenance & attribution

The training data combines human-authored and AI-generated material:

Source text & system prompt — written by kalaboster. Records derive from kalaboster's first-draft novel For The Small Shield (https://github.com/wordsum/For_The_Small_Shield), and the character/system prompt was authored by kalaboster.
input / output QA pairs — generated by Claude (Anthropic) from that first-draft novel.

In short: kalaboster wrote the novel and the system prompt; Claude turned the novel into the QA pairs.

Intended use & limitations

Creative role-play and flash-fiction storytelling in the Rise&Set universe. It stays in character as Carlos and is not a general-purpose assistant; it will frame answers as in-world stories and may invent lore. Not suitable for factual, safety-critical, or real-world advice.

License

Built with Meta Llama 3.

This model is a derivative of Meta Llama 3, so the Meta Llama 3 Community License governs its use, reproduction, distribution, and modification, and applies to its outputs. A full copy of the agreement ships with this repository:

LICENSE — Meta Llama 3 Community License Agreement
USE_POLICY.md — Meta Llama 3 Acceptable Use Policy
NOTICE — required attribution notice

Meta Llama 3 is licensed under the Meta Llama 3 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.

Terms published by Nous Research for the Hermes-2-Pro-Llama-3-8B base model also apply. If you redistribute this model or a derivative of it, carry these terms forward.

Citation

bibtex

@misc{llama-3-for-the-small-shield-v5-2026,
  title  = {Llama-3-for-the-small-shield (v5): Carlos the Barded Dwarf},
  author = {kalaboster},
  year   = {2026},
  url    = {https://huggingface.co/wordsum/for-the-small-shield}
}

This model card was drafted by Claude (Anthropic) on 2026-08-04 and updated on 2026-08-13. The model weights, tokenizer, and training data are the author's own work.