PratikBuilds/pocket-weather-theater-smollm2-135m-lora

Pocket Weather Theater SmolLM2 135M LoRA

This is a small LoRA adapter for Pocket Weather Theater, a Gradio-based creative toy built for the Build Small Hackathon. The adapter is tuned to produce short, strange, stage-like micro-scenes from compact inputs such as weather, prop, mood, and a hidden detail.

The goal is not general chat. The goal is a very specific voice: tiny theatrical scenes that feel playful, readable, and easy to share.

Model Details

• —Developed by: Pratik Shah
• —Base model: `HuggingFaceTB/SmolLM2-135M-Instruct`
• —Model type: PEFT LoRA adapter for causal language modeling
• —Language: English
• —License: Apache 2.0
• —Primary app: `build-small-hackathon/pocket-weather-theater`
• —Parameter scale: Tiny-model path using a 135M-parameter base model plus adapter weights

Intended Use

Use this adapter to generate compact creative scene text for Pocket Weather Theater-style prompts.

Good fits:

• —miniature scene writing
• —toy theatrical captions
• —prop dialogue
• —short whimsical performance snippets
• —local-first creative demos using a tiny model

Poor fits:

• —factual Q&A
• —coding help
• —medical, legal, financial, or safety-critical advice
• —long-form writing
• —unrestricted assistant chat

Prompt Shape

The model works best with structured, narrow prompts:

Write a tiny stage scene.
Weather: drizzle inside a desk drawer
Prop: teacup
Mood: delighted
Hidden detail: the teacup remembers the ocean

Return:
- title
- stage direction
- prop line
- closing image

Example Output Style

Title: Teacup Under Desk-Drizzle

The drawer opens like a very small sky. A teacup waits under the rain, bright with the patience of porcelain.

Teacup: I have held oceans smaller than this weather.

The drawer shuts softly. Somewhere inside, a saucer applauds.

Loading The Adapter

This is a LoRA adapter, so load it with the base model through PEFT:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "HuggingFaceTB/SmolLM2-135M-Instruct"
adapter_id = "PratikBuilds/pocket-weather-theater-smollm2-135m-lora"

tokenizer = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id, dtype="auto")
model = PeftModel.from_pretrained(base, adapter_id)

For the full app experience, use the linked Gradio Space.

Training Data

The adapter was trained on compact Pocket Weather Theater examples: structured prompt-to-scene pairs with weather, prop, mood, hidden detail, scene text, prop line, and shareable caption style.

The training data is intentionally narrow. This keeps the model focused on the app's voice instead of trying to become a general-purpose assistant.

Training Procedure

The adapter was fine-tuned with parameter-efficient LoRA training on top of SmolLM2-135M-Instruct. The training objective was next-token prediction over the structured creative examples.

High-level choices:

• —small base model for laptop-friendly inference
• —LoRA adapter instead of full fine-tune
• —short examples to match the final app interaction
• —low-temperature generation in the app for readable outputs

Evaluation

This model is evaluated primarily through the app experience rather than benchmark scores.

Review criteria used during development:

• —Does the output follow the requested structure?
• —Is the prop doing something meaningful in the scene?
• —Is the result short enough for the UI?
• —Is the voice playful without becoming unreadable?
• —Does it avoid generic assistant phrasing?

Limitations

• —The adapter is intentionally narrow and may perform poorly outside the Pocket Weather Theater format.
• —It can still generate repetitive, overly whimsical, or inconsistent text.
• —It should not be used for factual claims.
• —It may ignore formatting instructions if prompts are too broad.
• —It inherits limitations from the base model.

Safety Notes

The intended use is low-risk creative text generation. Users should still review outputs before publishing or sharing them. Do not use this model for advice, factual authority, or decisions affecting people.

Environmental Impact

This project uses a tiny 135M-parameter base model with a LoRA adapter. The small model choice was intentional: lower inference cost, lower hardware requirements, and a better fit for playful local-first experimentation.

Contact

Created by Pratik Shah.