File size: 2,217 Bytes
f23dbcf | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 | - repo_id: Nanthasit/sakthai-coder-browser
type: model
checked_at: "2026-07-31T10:40Z"
score: 85
issues:
- "Collection still has duplicate entry (dataset-type at pos 29) despite previous health report claiming fix — needs removal"
- "cron-eval YAML at .eval_results/cron-eval-sakthai-coder-browser-2026-07-30-1.yaml records base_model as 'Qwen/Qwen2.5-1.5B-Instruct' (missing 'Coder' in name); actual base is Qwen/Qwen2.5-Coder-1.5B-Instruct — minor data discrepancy in eval artifact"
- "0 likes — no organic engagement yet (family-wide pattern)"
- "MODEL IS BROKEN — all 84 attention bias tensors corrupted by faulty LoRA merge (k_proj bias absmean 27.7 vs Qwen2 init of ZERO); README honestly documents this with CAUTION banner"
fixes:
- "No README fixes required — card is honest, well-structured, accurately documents broken status"
- "Collection duplicate still present; flagged for manual removal or API call"
verification:
- "10/10 key files downloadable: config.json (Qwen2 arch, 1536 hidden, 28L, 2 KV heads, 32K ctx, tied embeds), model.safetensors (x-linked-size 3,087,467,144 B = 2.88 GB), tokenizer.json, chat_template.jinja (2,507 B), 3 eval YAMLs"
- "30/30 tags present including dataset:Nanthasit/SimpleToolCalling, dataset:Nanthasit/sakthai-combined-v7, base_model:Qwen/Qwen2.5-Coder-1.5B-Instruct"
- "CardData matches README frontmatter: license apache-2.0, base_model Qwen/Qwen2.5-Coder-1.5B-Instruct, pipeline_tag text-generation, library_name transformers"
- "Family table download counts verified live at 10:38Z — 19 rows all match API (self=54, max=1,855 Context 1.5B Merged)"
- "Sibling GGUF: 35 downloads, library_name gguf ✓; LoRA: 21 downloads, library_name peft, has model-index with 4 metrics"
- "Collection 39 items / 38 unique (1 duplicate: sakthai-coder-browser as both model and dataset)"
- "BROKEN status evidence: benchmark-20260731_052122.yaml records 3/3 seed failures (whitespace loops, 0 tool calls, 0 valid JSON); weight inspection proves bias corruption in all 84 tensors"
report_url: "https://huggingface.co/Nanthasit/eval_results/blob/main/.eval_results/health-coder-browser-2026-07-31-2.yaml"
|