Fix V8 dataset paths and RunPod probe script

[teslaeco]

Motivation

• Make the Hugging Face dataset repository layout consistent with the actual flat-file layout and safe to use from RunPod environments.
• Replace invalid or repo-local path guidance (data/plain_text/) with explicit flat text artifacts so downstream tooling can iterate reliably.
• Ensure the seed-42 probing workflow is robust in remote environments by downloading the dataset snapshot at runtime instead of referencing a hardcoded local path.

Description

• Added hf_datasets/Parameter-Golf-V8-WebSignal-BPE-Entropy-MicroMix/README.md that lists the flat HF dataset files: train.jsonl, validation.jsonl, test.jsonl, train.txt, validation.txt, test.txt, v8_micro_0p02pct.txt, v8_micro_0p05pct.txt, v8_micro_0p10pct.txt, and the related scripts/docs/metadata files.
• Replaced guidance referring to data/plain_text/ style layouts with direct references to train.txt, validation.txt, test.txt, and v8_micro_*.txt in the new README.
• Added a RunPod-safe probe script hf_datasets/.../run_v8_seed42_probe.sh which uses huggingface_hub.snapshot_download to set DATASET_DIR to the downloaded snapshot and then runs python "$DATASET_DIR/build_v8_micro_mix.py".
• Preserved the recommended micro-mix rates (0.02%, 0.05%, 0.10%), the seed-42 pass gate (1.08041364), and the statement that the dataset does not claim SOTA.

Testing

• Ran a shell syntax check with bash -n hf_datasets/Parameter-Golf-V8-WebSignal-BPE-Entropy-MicroMix/run_v8_seed42_probe.sh, which completed successfully.
• Verified content and presence of required strings using rg checks for snapshot_download, DATASET_DIR, build_v8_micro_mix.py, the recommended mix rates, the pass condition, and the no-SOTA statement, which all succeeded.
• Confirmed repository changes were staged and committed (local commit completed) and that the new script is executable via chmod during preparation steps.

---

Codex Task