SmolLM2-135M-Instruct-AO3

This model is a fine-tuned version of HuggingFaceTB/SmolLM2-135M-Instruct trained on the Fu01978/ao3_chat dataset.

It is designed to blend the instruction-following capabilities of SmolLM2 with the descriptive, narrative, and atmospheric prose styles commonly found in creative writing communities.

Model Description

• Model type: Causal Language Model (Fine-tuned)
• Language(s) (NLP): English
• License: Apache 2.0
• Finetuned from model: HuggingFaceTB/SmolLM2-135M-Instruct

Training Details

The model was fine-tuned for a short duration to "infuse" the base model with narrative flair without completely overwriting its general knowledge.

Training Hyperparameters

• Steps: 75
• Learning Rate: 5e-4
• Batch Size: 8
• Gradient Accumulation Steps: 2
• LR Scheduler Type: linear
• Precision: FP32 (Full Precision)

Loss Curve

The model showed a rapid descent in the first 20 steps, eventually stabilizing around a loss of 2.5-2.6.

Usage

This model uses the ChatML template. It is recommended to use the apply_chat_template method for best results.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Fu01978/SmolLM2-135M-Instruct-AO3"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float32, device_map="auto")

messages = [
    {"role": "system", "content": "You are a creative writing assistant."},
    {"role": "user", "content": "Write a scene about a rainy library."},
]

inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=150, do_sample=True, temperature=0.7)
print(tokenizer.decode(outputs[0][len(inputs[0]):], skip_special_tokens=True))

Intended Use & Limitations

• Roleplay & Fiction: Best suited for creative prompts and stylistic prose.
• Domain Shift: While it retains general knowledge, its explanations (e.g., science) may adopt a more narrative or "story-like" tone.