Hugging Face's logo

Datasets:
bestdive
/
details_bestdive__SmolLM3-3B-SFT-Free-Course

Modalities:
Text
Formats:
json
Size:
< 1K
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
Dataset Viewer
Auto-converted to Parquet Duplicate
results
string
model_name
string
num_samples
int64
{"lighteval|gsm8k|0": {"extractive_match": 0.4086429112964367, "extractive_match_stderr": 0.013540639733342422}, "all": {"extractive_match": 0.4086429112964367, "extractive_match_stderr": 0.013540639733342422}}
bestdive/SmolLM3-3B-SFT-Free-Course
1,319

Smol course SFT evaluation - Kay Zheng

Actual full GSM8K test evaluation of bestdive/SmolLM3-3B-SFT-Free-Course, adapter revision 0484e028b494d605a267050a949c9266edadd16b, merged with pinned SmolLM3-3B-Base before evaluation.

  • Full 1319 test examples, zero-shot, original extractive_match: 0.4086429112964367 (stderr 0.013540639733342422).
  • Free Google Colab T4, no paid HF Jobs; cost 0.
  • lighteval 0.11.0, vLLM 0.10.1.1, Transformers 4.57.1, Python 3.12.
  • Dataset-address correction from malformed lighteval/openai/gsm8k to openai/gsm8k; audit included.
  • xxhash 4.0.1 rejected string hashing in detail logging after all responses had been generated. Pinned xxhash 3.5.0 and reused all 1319 cached responses for successful scoring; both logs included. Questions and scoring were unchanged.
  • raw-results.json and gsm8k-details.parquet are original evaluator outputs. results.jsonl packages the same result mapping for the course leaderboard's results/latest interface.
  • Training, evaluation and submission prepared with AI assistance. Official review pending; this dataset is not a course certificate.
Downloads last month
8
Number of rows:
1
Total file size:
6.82 MB