DSTI/traffic-accidents-reports-kd-smollm2-360M-7k

Accident Reporting KD Dataset (One-Paragraph)

Short description.
A training/evaluation dataset for generating one-paragraph accident/incident reports from structured facts.
This dataset mixes gold human targets from zBotta/traffic-accidents-reports-5k with teacher-generated soft targets produced by the model zBotta/smollm2-accident-reporter-360m to support knowledge distillation (KD) of a smaller student.

Output style: a single paragraph, neutral tone, covering What, When, Where, Who, How, Why, ContingencyActions.
Language: English.

---

Contents

• train.jsonl — training split.
  Contains all original gold examples plus a KD augmentation subset where the target is generated by the teacher.
• eval.jsonl — evaluation split (kept as in the source dataset; gold references only).

Each line is a JSON record.

---

Data schema

Minimal required fields:

Field	Type	Description
input	string	Structured facts as a flat string: What: …, When: …, Where: …, Who: …, How: …, Why: …, ContingencyActions: …
target	string	One single paragraph narrative generated from the facts (gold or teacher-generated in train; gold in eval).
source	string	"gold" or "teacher" for training rows.

---

Provenance & generation process

Base dataset: zBotta/traffic-accidents-reports-5k (splits train/eval, fields input/target).
Teacher model for KD: zBotta/smollm2-accident-reporter-360m.

KD augmentation:

• Randomly select a fraction of training inputs (used 50%).
• Prompt the teacher deterministically (no sampling) using the same instruction template used for SFT.
• Decoding parameters (best mean ce_sim score for teacher):
  • do_sample=True
  • max_new_tokens=300
  • temperature=0.7
  • top_p=0.6
  • repetition_penalty=1
  • eos_token_id and pad_token_id set to tokenizer defaults
• Post-processing:
  • Keep only the text after the response marker (if used).
  • Truncate at the first double newline or section marker (###, JSON, code fences).
  • Collapse whitespace to a single line; remove repeated token artifacts (e.g., “word word word”).
  • Ensure one paragraph only (no line breaks).
• Merge augmented rows with the original gold training data.

The eval split is not altered and remains gold-only for unbiased evaluation.

---

How to load

from datasets import load_dataset

ds = load_dataset("zBotta/accident-reporting-kd", split=None)  # replace with your repo id
print(ds["train"][0].keys())   # -> dict_keys(['input','target', ...])

# Iterate examples
for ex in ds["train"].select(range(3)):
    print(ex["input"][:120], "...")
    print(ex["target"])
    print("---")

Citation

If you use this dataset, please cite:

The source dataset: DSTI/traffic-accidents-reports-5k

The teacher model: DSTI/smollm2-accident-reporter-360m

Relevant distillation literature, e.g.: Hinton, Vinyals, Dean. Distilling the Knowledge in a Neural Network (2015).

@misc{accident_kd_dataset,
  title  = {Accident Reporting KD Dataset (One-Paragraph)},
  author = {zBotta, SamdGuizani},
  year   = {2025},
  note   = {Synthetic KD augmentations generated by zBotta/smollm2-accident-reporter-360m}
}