qwen3-5-0-8b-rmsnorm-experiment

# Qwen3.5-0.8B RMSNorm Experiment Dataset This dataset captures every observed Hugging Face Jobs run for the `qwen3_5_0_8b_rmsnorm` benchmark family. ## Files - `runs.parquet`: one row per HF Jobs attempt with run-level metrics and derived classifications. - `microbenchmarks.parquet`: one row per shape per completed run. - `run_failures.parquet`: normalized failure causes for the non-successful attempts. - `experiment.duckdb`: a portable DuckDB database with the same tables plus convenience views. - `raw_logs/`: the original `hf jobs logs` output for each job id. - `reports/`: extracted benchmark reports for completed runs. - `analysis_summary.md`: a concise markdown report generated from SQL over the DuckDB database. - `sql_examples.sql`: example queries used in the report and the HF Space. ## Derived fields - `classification=semantics_correct`: the run passed correctness and its patched transformer output stayed within the benchmark tolerance. - `classification=fast_but_semantically_incorrect`: the run completed and benchmarked well, but the patched transformer path drifted too far from the baseline. - `failure_type`: a normalized category for non-successful attempts. ## Row counts - runs: 6 - completed runs: 3 - semantics-correct runs: 1