ykumards

smollm2-135m-bedtime-stories

README

License: apache-2.0

Prompt

It was tuned on exactly this user turn, with no system turn. Start the reply with Once upon a time, there was a little: without that opening it tends to ramble in the second person.

markdown
<|im_start|>user
Here is something from my notebook:

{text}

Tell me a short bedtime story about it.<|im_end|>
<|im_start|>assistant
Once upon a time, there was a little

The story ends with <|im_end|>. Sample at temperature 0.5, top-p 0.9; keep the notebook text under about 120 words.

python
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "ykumards/smollm2-135m-bedtime-stories"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo)

text = "Took the kids to the zoo. The penguins were the best part."
opening = "Once upon a time, there was a little"
request = f"Here is something from my notebook:\n\n{text}\n\nTell me a short bedtime story about it."
prompt = tokenizer.apply_chat_template([{"role": "user", "content": request}],
                                       tokenize=False, add_generation_prompt=True) + opening
ids = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).input_ids
out = model.generate(ids, max_new_tokens=300, do_sample=True, temperature=0.5, top_p=0.9)
print(opening + tokenizer.decode(out[0, ids.shape[1]:], skip_special_tokens=True))

The bundled chat template adds no system prompt (SmolLM2's original inserts a default one, which this model never saw), so apply_chat_template builds the training prompt; append the opening yourself.

On a reMarkable 2

unremarkable/story-q8.bin and unremarkable/story-q8.tok are the same weights in the unremarkable engine's Q8_0 format. On the tablet's two Cortex-A7 cores it takes about 5 s to the first word and writes about 6.4 tokens/s: roughly 30 s per story, in 171 MiB of memory.

Training

Data: 56,910 synthetic pairs from Qwen3.5-9B (AWQ, on vLLM). Each pair is a piece of notebook text written from a random person, mood and day (a single word, a phrase, a list, a note, a diary entry or a copied line) with its key thing, and a story of about 70 to 120 words written from the text alone: built around the key, a small plot that works out, simple words, no stock mood words, and nobody has to fall asleep at the end.
Filtering: pairs whose story lost the key, leaned on words like soft, warm and cozy, mentioned the notebook, or ran outside 60 to 180 words were dropped, leaving 53,663 (94%). 15% of prompts got handwriting-style typos.
Recipe: one epoch over two copies of the pairs (105k examples), loss on the story only, AdamW at 1e-4 with cosine decay, bf16 autocast, batches of 8,192 tokens, on one RTX 4090.

Code: train/.

Evaluation

24 hand-written selections (never generated), two seeds each, at temperature 0.8 through the Q8 engine, without the opening:

Table with columns: model, kept the key, ended asleep, mood words /100, notebook leaks
model	kept the key	ended asleep	mood words /100	notebook leaks
SmolLM2-135M-Instruct	32/48	8/48	1.2	2
an earlier diary fine-tune	38/48	46/48	9.7	3
this model	48/48	3/48	0.2	0

Limitations

It is a 135M model. Stories stay on topic but the middle can wander, and it slips: penguins with orange feathers, "Miss grandpa today" read as a name.
It does not get jokes or sarcasm.
For hard entries ("Mom is in hospital. Scared.") it stays gentle but cannot really comfort. It is not a substitute for talking to someone.
English only. The training data is synthetic.

License

Apache 2.0, like the base model. The training stories were generated with Qwen3.5-9B.

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

ykumards

Model Tree

Base

HuggingFaceTB/SmolLM2-135M-Instruct

Fine-tuned

this model

Input Modalities

Text

Output Modalities