---

license: apache-2.0

language:

- en

library_name: gguf

pipeline_tag: text-generation

tags:

- gguf

- quantized

- llama-cpp

- ollama

- lm-studio

- sft

- distillation

- fable

- creative-writing

base_model: ermiaazarkhalili/Granite-4.1-3B-SFT-Fable5-Glint

---

Granite-4.1-3B-SFT-Fable5-Glint — GGUF

Quantized GGUF builds of Granite-4.1-3B-SFT-Fable5-Glint, a

ibm-granite/granite-4.1-3b model supervised-fine-tuned on the

FABLE-5 trace corpus. These files run locally with

llama.cpp, Ollama,

LM Studio, and any GGUF-compatible runtime — no GPU required for the

smaller quants.

Overview

| | |

|---|---|

| Fine-tuned model | Granite-4.1-3B-SFT-Fable5-Glint |

| Base model | ibm-granite/granite-4.1-3b |

| Parameter class | 3B |

| Model family | dense |

| Training method | LoRA SFT (distillation) |

| Domain | FABLE-5 creative / agentic traces |

| Format | GGUF (this repo) · safetensors (merged repo) |

Standard dense decoder base.

What is FABLE-5?

This model was fine-tuned on FABLE-5-Glint, a cleaned corpus of FABLE-5 pi-agent

reasoning traces (each target completion may include a <think>…</think> reasoning span

followed by the response). Training used assistant-only loss masking so the model learns to

produce the response, not echo the prompt. The dataset is private; the fine-tuned weights

are public.

Available Quantizations

| File | Quant | Size | Notes |

|------|-------|------|-------|

| granite-4.1-3b-sft-fable5-glint.q4_k_m.gguf | Q4_K_M | ~2.0 GB | Recommended — best quality/size balance |

| granite-4.1-3b-sft-fable5-glint.q5_k_m.gguf | Q5_K_M | ~2.3 GB | Higher quality |

| granite-4.1-3b-sft-fable5-glint.q8_0.gguf | Q8_0 | ~3.4 GB | Maximum quality (near-lossless) |

Which to pick: Q4_K_M is the best size/quality trade-off for most users. Use Q5_K_M

if you have spare RAM/VRAM and want a little more fidelity, or Q8_0 for near-lossless output

when size is not a concern.

Usage

Ollama

# Pull + run the recommended Q4_K_M quant directly from the Hub
ollama run hf.co/ermiaazarkhalili/Granite-4.1-3B-SFT-Fable5-Glint-GGUF:Q4_K_M "Write a short story about a clockwork fox."

To pin a different quant, swap the tag (e.g. :Q5_K_M, :Q8_0).

llama.cpp

# Download a single quant, then run it
huggingface-cli download ermiaazarkhalili/Granite-4.1-3B-SFT-Fable5-Glint-GGUF granite-4.1-3b-sft-fable5-glint.q4_k_m.gguf --local-dir .
llama-cli -m granite-4.1-3b-sft-fable5-glint.q4_k_m.gguf -p "Write a short story about a clockwork fox." -n 512

# Or serve an OpenAI-compatible endpoint
llama-server -m granite-4.1-3b-sft-fable5-glint.q4_k_m.gguf --host 0.0.0.0 --port 8080

LM Studio

Search for ermiaazarkhalili/Granite-4.1-3B-SFT-Fable5-Glint-GGUF in LM Studio, or download a .gguf above and load it from disk.

Python (llama-cpp-python)

from llama_cpp import Llama
llm = Llama.from_pretrained(repo_id="ermiaazarkhalili/Granite-4.1-3B-SFT-Fable5-Glint-GGUF", filename="*q4_k_m.gguf", n_ctx=4096)
out = llm.create_chat_completion(
    messages=[{"role": "user", "content": "Write a short story about a clockwork fox."}]
)
print(out["choices"][0]["message"]["content"])

Prompt format

Use the base model's chat template (applied automatically by Ollama / LM Studio /

create_chat_completion). The model was trained on 2-turn user → assistant chats. For

thinking-style bases, the model may emit a <think>…</think> span before its answer.

Training details

| Hyperparameter | Value |

|---|---|

| Method | LoRA SFT, merged to 16-bit then quantized |

| LoRA rank / α | 16 / 16 |

| Learning rate | 0.0002 |

| LR scheduler | cosine (warmup 0.03) |

| Max sequence length | 4096 |

| Epochs | 3 |

| Loss masking | assistant-only |

| Quantization toolchain | llama.cpp convert_hf_to_gguf + llama-quantize |

A deterministic ~5% slice of the corpus was held out from training for evaluation.

Intended use & limitations

• Intended: local/offline creative writing, reasoning-trace style generation, and

experimentation with FABLE-5-distilled behavior on consumer hardware.

• Limitations: inherits the base model's knowledge cutoff and biases; quantization

(especially Q4_K_M) trades some fidelity for size; not safety-tuned for production use

without additional guardrails. Outputs may be fictional/unverified.

Citation

@misc{azarkhalili2026granite413bsftfable5glint,
    author = {Azarkhalili, Behrooz},
    title  = {Granite-4.1-3B-SFT-Fable5-Glint: FABLE-5 SFT distillation (GGUF)},
    year   = {2026},
    publisher = {Hugging Face},
    url    = {https://huggingface.co/ermiaazarkhalili/Granite-4.1-3B-SFT-Fable5-Glint-GGUF}
}