SmolLM2-accident-reporter-1.7B

DSTI/SmolLM2-accident-reporter-1.7B · Accident Reporter

Model Details

SmolLM2-accident-reporter-1.7B is a LoRA fine-tuned variant of SmolLM2-1.7B-Instruct, trained for knowledge-distilled accident reporting. The model generates concise, neutral one-paragraph traffic accident/incident reports from structured facts. Each output is designed to cover the key aspects of an event: What, When, Where, Who, How, Why, and Contingency Actions. This model is intended for tasks in structured event-to-text generation, summarization of incidents, and training student models with KD signals.

Model Description

• Base Model: unsloth/SmolLM2-1.7B-Instruct
• Language(s) (NLP): English
• License: apache-2.0
• Dataset: DSTI/traffic-accidents-reports-kd-smollm2-360M-7k

Uses

Direct Use

• Automatic generation of one-paragraph traffic accident reports.
• Knowledge distillation research for event-to-text tasks.
• Supporting structured-to-freeform NLP generation benchmarks.

Bias, Risks, and Limitations

• The model follows a neutral reporting tone, but may omit minor details not emphasized in training.
• Not suitable for real-time or legal use cases without human verification.
• Performance is limited to traffic/incident report style and may not generalize to unrelated domains.

How to Get Started with the Model

Use the code below to get started with the model.

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel


tokenizer = AutoTokenizer.from_pretrained("unsloth/SmolLM2-1.7B-Instruct",)
base_model = AutoModelForCausalLM.from_pretrained(
    "unsloth/unsloth/SmolLM2-1.7B-Instruct",
    device_map={"": 0}
)

model = PeftModel.from_pretrained(base_model,"DSTI/SmolLM2-accident-reporter-1.7B")


question ="""
What: Minor collision between Vehicle A (compact car) and Vehicle B (minivan) in parking lot
When: Occurrence: December 9, 2025, 13:00; Discovery: December 9, 2025, 13:01
Where: Shopping center parking lot, Lot B
Who: Ms. Karen Liu – compact car driver (Vehicle A), Mr. Thomas Barnes – minivan driver (Vehicle B)
How: Vehicle A misjudged turning space and touched Vehicle B
Why: Driver inattention during parking maneuver
ContingencyActions: Drivers exchanged info, no injuries reported, security cameras documented accident
"""

messages = [
    {"role" : "user", "content" : question}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize = False,
    add_generation_prompt = True, 
)

from transformers import TextStreamer
_ = model.generate(
    **tokenizer(text, return_tensors = "pt").to("cuda"),
    max_new_tokens = 512,
    streamer = TextStreamer(tokenizer, skip_prompt = True),
)

For pipeline:

from transformers import pipeline, AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

tokenizer = AutoTokenizer.from_pretrained("unsloth/SmolLM2-1.7B-Instruct")
base_model = AutoModelForCausalLM.from_pretrained("unsloth/SmolLM2-1.7B-Instruct")

model = PeftModel.from_pretrained(base_model, "DSTI/SmolLM2-accident-reporter-1.7B")

question ="""
What: Minor collision between Vehicle A (compact car) and Vehicle B (minivan) in parking lot
When: Occurrence: December 9, 2025, 13:00; Discovery: December 9, 2025, 13:01
Where: Shopping center parking lot, Lot B
Who: Ms. Karen Liu – compact car driver (Vehicle A), Mr. Thomas Barnes – minivan driver (Vehicle B)
How: Vehicle A misjudged turning space and touched Vehicle B
Why: Driver inattention during parking maneuver
ContingencyActions: Drivers exchanged info, no injuries reported, security cameras documented accident
"""

pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
messages = [
    {"role": "user", "content": question}
]
pipe(messages)

Training Details

Accident Reporting KD Dataset (One-Paragraph)

The model was fine-tuned on the Accident Reporting KD Dataset, which consists of:

• Gold human-written targets from zBotta/traffic-accidents-reports-5k
• Teacher-generated reports from zBotta/smollm2-accident-reporter-360m, providing soft targets for knowledge distillation (KD).

Dataset size: ~6K samples.

Language: English.

Result

• Training Loss: 2.43 >> 0.70
• Eval Loss: 2.41 >> 0.71

Citation

If you use this model, please cite:

The source dataset: DSTI/traffic-accidents-reports-kd-smollm2-360M-7k

@misc{SmolLM2-accident-reporter-1.7B,
  title  = {Accident Reporting model (One-Paragraph)},
  author = {Rustam Shiriyev},
  year   = {2025}
}

Framework versions

• PEFT 0.15.2