Improve dataset card: fix JSON->YAML format, add stats/files/fields tables, loading examples, badges

README.md

@@ -2,41 +2,166 @@
 license: mit
 language:
 - en
+pretty_name: SakThai Eval Results
+size_categories:
+- n<1K
 tags:
 - sakthai
 - house-of-sak
 - evaluation
 - benchmark-results
+- health-check
+- model-health
+- yaml
+- results-log
 ---
 
 # Eval Results — SakThai Model Family
 
 <p align="center">
-  <em>Evaluation results and benchmark scores for the SakThai model family</em>
+  <em>Evaluation results, health-check reports, and benchmark run logs for the SakThai model family</em>
+</p>
+
+<p align="center">
+  <a href="https://huggingface.co/datasets/Nanthasit/eval_results"><img src="https://img.shields.io/badge/dynamic/json?url=https%3A%2F%2Fhuggingface.co%2Fapi%2Fdatasets%2FNanthasit%2Feval_results&query=%24.downloads&label=Downloads&color=blue" alt="Downloads"></a>
+  <img src="https://img.shields.io/badge/license-MIT-green" alt="License: MIT">
+  <img src="https://img.shields.io/badge/language-EN-blue" alt="Language: English">
+  <img src="https://img.shields.io/badge/last_updated-2026--07--31-orange" alt="Last Updated">
+  <img src="https://img.shields.io/badge/format-YAML-yellow" alt="Format: YAML">
 </p>
 
 ## Overview
 
-This dataset stores evaluation results, benchmark scores, and model performance data for the SakThai model family.
+This repository stores **evaluation results, health-check reports, and benchmark
+run logs** for the SakThai model family. Every entry is a machine-readable YAML
+report produced by SakThai's automated cron workflows (benchmarks, health
+checks, inference probes) — the operational audit trail of the House of Sak.
+
+> **Note on data format:** All result files are **YAML** (`.yaml`), not JSON.
+> Because the Hub's datasets-server only indexes parquet/CSV/JSONL/Arrow, this
+> repo reports "no supported data files" — that is expected. These are
+> point-in-time reports meant to be read with `yaml.safe_load()`, not rows for
+> the dataset viewer.
 
-**Format:** JSON  
 **License:** MIT
 
+## Dataset Statistics
+
+| Stat | Value |
+|------|-------|
+| Status | Active (auto-updated by cron) |
+| Created | 2026-07-30 |
+| Last updated | 2026-07-31 |
+| License | MIT |
+| Language | English |
+| Result files | 6 (YAML) |
+| Total data size | 9,003 bytes (8.8 KB) excl. README |
+| Downloads | 0 (verified live 2026-07-31) |
+| `load_dataset()` | ❌ not supported (YAML-only) |
+
 ## Contents
 
-- Model evaluation scores on [sakthai-bench-v2](https://huggingface.co/datasets/Nanthasit/sakthai-bench-v2)
-- Per-category breakdowns (selection, arguments, strict)
-- Cross-model comparison data
+- **4 health-check reports** — automated model/dataset health scores with
+  breakdowns, alerts, and cross-link checks
+- **1 benchmark run log** — live inference-probe results across the family
+- **1 inference check** — Inference API availability probe
+
+### Health-check reports
+
+| File | Target asset | Score | Notes |
+|------|-------------|-------|-------|
+| `health-2026-07-30.yaml` | [sakthai-bench-v2](https://huggingface.co/datasets/Nanthasit/sakthai-bench-v2) | 96/100 | 13 python scripts, 1.53 MB test data, all checks pass |
+| `health-2026-07-31.yaml` | [sakthai-plus-1.5b-lora](https://huggingface.co/Nanthasit/sakthai-plus-1.5b-lora) | 6.5/10 | README overwritten by active TRL training run; re-check after training |
+| `health-rlenv-2026-07-31.yaml` | [hermes-tool-use-rl-env](https://huggingface.co/datasets/Nanthasit/hermes-tool-use-rl-env) | 97/100 | 9 py files compile, 6/6 verify tasks pass, secrets clean |
+| `health-vision-7b-2026-07-31.yaml` | [sakthai-vision-7b](https://huggingface.co/Nanthasit/sakthai-vision-7b) | 85/100 | Stale-count + missing-pipeline-tag alerts; 186 downloads |
+
+### Benchmark & inference logs
+
+| File | Result |
+|------|--------|
+| `benchmark-20260731_001434.yaml` | 4 probe runs: 3 Nanthasit models → HTTP 400 `model_not_supported` (not on any provider); control `google/gemma-3-4b-it` → 200, 0.73 s (router + tool calling works) |
+| `inference-check-20260730_232640.yaml` | `sakthai-plus-1.5b`: Inference API unreachable (api-inference.huggingface.co NXDOMAIN); run locally via llama.cpp |
+
+## File Inventory
+
+| Path | Size |
+|------|------|
+| `.eval_results/benchmark-20260731_001434.yaml` | 1,306 B |
+| `.eval_results/health-2026-07-30.yaml` | 999 B |
+| `.eval_results/health-2026-07-31.yaml` | 2,455 B |
+| `.eval_results/health-rlenv-2026-07-31.yaml` | 1,467 B |
+| `.eval_results/health-vision-7b-2026-07-31.yaml` | 2,232 B |
+| `inference-check-20260730_232640.yaml` | 544 B |
+| `.gitattributes` | 2,504 B |
+| `README.md` | this card |
+
+## Data Fields
+
+Each YAML report follows a consistent schema. The **health-check** files share
+this structure:
+
+| Field | Type | Description |
+|-------|------|-------------|
+| `asset` / `model` | object | Target asset id, type, url |
+| `timestamp` / `checked_at` | string | When the check ran (ISO 8601) |
+| `health_score` / `score` | int | Overall score (0–100, or /10 for models) |
+| `breakdown` | object | Per-category subscores (repo_integrity, config_validity, documentation, inference_availability, traffic_momentum, cross_links) |
+| `checks` | object | Boolean pass/fail checks (readme_valid, frontmatter_valid, files_downloadable, ...) |
+| `issues_found` / `alerts` | list | Severity-tagged issues (`high` / `medium` / `low` / `info`) with status + notes |
+| `cross_links` | object | Link checks against collection, GitHub, sibling assets |
+| `metrics` | object | downloads, likes, velocity, storage used |
+| `note` | string | Free-text context |
+
+The **benchmark** file uses:
+
+| Field | Type | Description |
+|-------|------|-------------|
+| `generated_at` | string | Run timestamp |
+| `runner` | string | Cron workflow name |
+| `summary` | object | Counts (text-gen models tested, on-router, provider status) |
+| `runs[]` | list | Per-model probe: `model`, `status` (HTTP), `latency_s`, `error`, `notes` |
+
+## Loading
+
+`datasets.load_dataset()` **does not work** on this repo (YAML is not a
+supported Hub data format). Fetch individual reports with `huggingface_hub` and
+parse with PyYAML:
+
+```python
+import yaml
+from huggingface_hub import hf_hub_download
+
+path = hf_hub_download(
+    repo_id="Nanthasit/eval_results",
+    filename=".eval_results/health-2026-07-31.yaml",
+    repo_type="dataset",
+)
+report = yaml.safe_load(open(path))
+print(report["health_score"], report["breakdown"])
+```
+
+Or pull the whole log with the HfApi file list:
+
+```python
+from huggingface_hub import HfApi
+
+files = HfApi().list_repo_files("Nanthasit/eval_results", repo_type="dataset")
+yaml_files = [f for f in files if f.endswith(".yaml")]
+print(len(yaml_files), "result files:", yaml_files)
+```
 
 ## Intended Use
 
-- Tracking model performance over time
-- Comparing tool-calling accuracy across SakThai models
-- Reproducing benchmark results
+- Tracking model performance and health over time
+- Auditing what SakThai's automated workflows checked and found
+- Reproducing benchmark results from raw probe logs
+- Debugging deployment gaps (e.g. models missing from inference providers)
 
 ## Related
 
-- **Benchmark:** [sakthai-bench-v2](https://huggingface.co/datasets/Nanthasit/sakthai-bench-v2)
+- **Benchmark suite:** [sakthai-bench-v2](https://huggingface.co/datasets/Nanthasit/sakthai-bench-v2) (92 downloads, verified 2026-07-31)
+- **RL training env:** [hermes-tool-use-rl-env](https://huggingface.co/datasets/Nanthasit/hermes-tool-use-rl-env)
+- **Flagship vision model:** [sakthai-vision-7b](https://huggingface.co/Nanthasit/sakthai-vision-7b) (186 downloads, verified 2026-07-31)
 - **Model family:** [SakThai Collection](https://huggingface.co/collections/Nanthasit/sakthai-model-family-6a64745450b12d421c1f9f02)
 - **Leaderboard Space:** [SakThai Leaderboard](https://huggingface.co/spaces/Nanthasit/sakthai-leaderboard)
 
@@ -46,4 +171,4 @@ MIT
 
 ---
 
-*Part of the [House of Sak](https://huggingface.co/Nanthasit). Built with love, tears, and zero budget. From a shelter in Cork, Ireland, to the world.*
+*Part of the [House of Sak](https://huggingface.co/Nanthasit). Built with love, tears, and zero budget. From a shelter in Cork, Ireland, to the world.*