ykumards/smollm2-135m-bedtime-stories
SmolLM2-135M bedtime stories
Give it something from your notebook, whether a word, a to-do list or a few lines about your day, and it tells a short, gentle bedtime story about it. A fine-tune of SmolLM2-135M-Instruct made for unremarkable, which runs it on a reMarkable 2: circle some handwriting, tap a button, and the story is written on the tablet itself.
Notebook: Hmm, woke up on the wrong side of the bed again. How does this keep happening? Always on the right! Subconscious, do better! Story: Once upon a time, there was a little frog named Emil. He woke up on the wrong side of the bed again. His legs felt wobbly. He wanted to jump high to see the big green pond. But the water was too deep for him. Emil did not give up… He landed right on the edge of the pond… He is happy and settled.
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.
<|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 littleThe story ends with <|im_end|>. Sample at temperature 0.5, top-p 0.9; keep the notebook text under about 120 words.
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:
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.