neonforestmist

smolgpt-fables

README

License: apache-2.0

What it does

Writes short English stories and fables with 1–6 numbered scenes.
Accepts separate Name and About them fields for each main character.
Tries to preserve requested names, setting details, important objects, and endings.
Runs with Transformers, llama.cpp-compatible apps, or Core ML on Apple devices.

See it in action

Open SmolGPT-Fables Studio, choose the model and number of scenes, then describe each character in their own card. The Studio returns a clean story preview and a Markdown download.

Try a story

The easiest way to use the model is SmolGPT-Fables Studio. You do not need to write Markdown or learn a prompt format.

text
Story idea: A paper moth leads a mapmaker to a neighborhood that vanished from every map.
Genre: Cozy fantasy
Scenes: 4

Name: Mina Vale
About them: A careful apprentice mapmaker who wants credit for her work.

Name: Orin Reed
About them: A retired courier who remembers roads that no longer appear on maps.

Setting: A rain-bright canal city whose maps change at midnight.
Ending: Mina restores the neighborhood without exposing the people who protected it.

The expected output looks like this:

markdown
### Scene 01: The Map That Moved

Mina Vale's brass compass turned away from north and followed a paper moth
through the archive window.

### Scene 02: The Unwritten Road

Orin Reed recognized the vanished towpath while Mina checked each bridge
against the changing map.

This short excerpt shows the expected format. It is not an evaluated generation or a promise that every response will use these words.

Use it with Transformers

python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "neonforestmist/smolgpt-fables"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="auto",
)

messages = [
    {
        "role": "user",
        "content": (
            "Write a two-scene cozy fable about Mina Vale, a careful mapmaker, "
            "and Orin Reed, a retired courier. Set it in a canal city at "
            "midnight and end with Mina restoring a vanished neighborhood."
        ),
    }
]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, do_sample=False, max_new_tokens=700)
new_tokens = outputs[0, inputs["input_ids"].shape[1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))

Downloads

Table with columns: Pick, File, Size, Best for
Pick	File	Size	Best for
Standard	`model.safetensors`	3.42 GB	Transformers and full model precision
Compact	`SmolGPT-Fables-v1-Q4_K_M.gguf`	1.06 GB	Runtimes that support `Q4_K_M` GGUF models
Core ML	`coreml/SmolGPT-Fables-v1-CoreML-INT4.mlpackage`	0.96 GB

The evaluation results below apply to the standard Transformers model. The GGUF version has not been evaluated separately. The Core ML package passed a stateful prediction smoke test and matched the FP16 model's top token.

Use it in an Apple app

The Core ML package is stored in this repository so a SwiftUI project can stay small. An app can download the package once, verify it with coreml/coreml-checksums.sha256, compile it on the device, and cache it for offline stories.

The package uses INT4 weights, a 2,048-token context, and a stateful key/value cache. Its minimum deployment versions are iOS 18 and macOS 15.

How well it works

The Studio-focused check is strong on the behavior this interface needs:

Table with columns: Check, Result
Check	Result
Stories containing every requested full name	60 / 60
First name combined with the wrong surname	0
Studio behavior checks	All passed

A stricter writing-quality suite passed every check on 32 of 60 stories. Its lowest measured document-diversity score was 0.839416, just below the required 0.850000. The most common misses were incomplete causal resolution and scene length. These measurements are useful signals, not guarantees for every story.

Known limits

Names and requested scene counts are learned behavior, not hard guarantees.
Longer stories can lose continuity or end weakly.
The model can repeat phrases or flatten cultural nuance.
It is English-focused and intended for short stories rather than factual advice.
Review and edit generated writing before sharing it.

Training data

SmolGPT-Fables v1 was adapted from HuggingFaceTB/SmolLM2-1.7B-Instruct using the published neonforestmist/smolgpt-markdown-stories dataset. The synthetic and curated stories cover varied genres, settings, character descriptions, required details, and endings.

The published BF16 artifact contains 1,711,376,384 parameters across 218 tensors. It uses micro update 8 over adapter update 68 and the smolgpt-fables-smollm3-chat-v5 prompt contract.

Table with columns: Artifact, SHA-256
Artifact	SHA-256
`model.safetensors`	`00d2c2b4b01ce0a13fbec5ba79f4fb843a751af2780e939ab431a092f1789207`
`training_manifest.json`	`22b90db42797c891bb69ac526ff237ef973be05827722ae83d2e3fdb794702ea`
`SmolGPT-Fables-v1-Q4_K_M.gguf`	`6d4d38e9075bcae3cfe59ed37c0acc3425ad31296de7530611e63628424a7c35`
GGUF quantization manifest	`87da452fd6df5658b8e9f07844a84993f656ae1b3f580e560eba018a37cc5a2f`
GGUF checksums file

The GGUF file contains 218 tensors and is 1,055,609,376 bytes. It was converted with llama.cpp commit aff6eb6e7503538fec1532dec2f584bc7a4a4e4d.

Strict-suite diagnostic occurrences:

Table with columns: Diagnostic, Count
Diagnostic	Count
Incomplete causal resolution	17
Scene length	16
Character-role retention	1
Required detail	2
Scene heading	2
Unresolved ending	1

Character-name audit report SHA-256: 38a1ba15d6ae8545ae71a9953077d0626aba5c85ccf7f3844514bba04188f3a6.

License

Apache-2.0. Review the SmolLM2-1.7B-Instruct model card alongside this one when assessing intended use and limitations.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

neonforestmist

Model Tree

Base

HuggingFaceTB/SmolLM2-1.7B-Instruct

Quantized

this model

Input Modalities

Text

Output Modalities