model: Nanthasit/sakthai-coder-browser
eval_date: "2026-07-31T11:37Z"
eval_type: health_check

metadata:
  pipeline_tag: text-generation
  library_name: transformers
  license: apache-2.0
  base_model: Qwen/Qwen2.5-Coder-1.5B-Instruct
  created: "2026-07-30T19:21Z"
  last_modified: "2026-07-31T10:50Z"

core_metrics:
  downloads: 54
  likes: 0
  age_days: 0.7
  downloads_per_day: 79.6
  siblings: 9
  total_repo_bytes: 3098911420

card_content:
  readme_size_bytes: 15959
  frontmatter_valid: true
  cross_links:
    total: 30
    working: 30
    broken: 0
  shields_badges: 7
  family_table_rows: 19
  family_table_download_accuracy: all_match

tree_cleanliness:
  siblings: 9
  top_level_entries: 9
  stray_dev_files: false
  stale_lfs_entries: false
  note: "Clean tree — no stray files"

config_verification:
  architecture: Qwen2ForCausalLM
  hidden_size: 1536
  num_layers: 28
  num_attention_heads: 12
  num_kv_heads: 2
  max_position: 32768
  vocab_size: 151936
  torch_dtype: bfloat16
  model_type: qwen2

eval_yamls:
  - benchmark-20260731_052122.yaml
  all_parse: true
  model_verdict: MODEL_BROKEN_BIAS_CORRUPTION
  benchmark_valid: false

health_score:
  overall: 65.0
  max: 100
  components:
    popularity:
      score: 18.0
      weight: 0.20
      input: "54 downloads in 0.7 days"
    momentum:
      score: 79.6
      weight: 0.20
      input: "79.6 dl/day"
    benchmarks:
      score: 50
      weight: 0.25
      input: "eval YAML exists (benchmark results: MODEL_BROKEN_BIAS_CORRUPTION), no model-index"
    card_quality:
      score: 90
      weight: 0.20
      input: "frontmatter valid, family table accurate, broken-weight warning present"
    repo_hygiene:
      score: 100
      weight: 0.15
      input: "clean tree, no junk, standard .gitattributes"

assessment:
  health: MODERATE
  summary: >-
    Model is accurately documented as BROKEN (bias corruption from faulty LoRA merge).
    All cross-links work (23/23 HF + shields), frontmatter is valid, config matches card claims,
    family table download counts are live-accurate (verified against API), and the repo tree is clean.
    Health score limited by the broken-weights verdict and lack of a model-index.
    The honest BROKEN warning in the README is the right call — this is a quality-of-documentation
    success even though the model itself isn't deployable.

issues:
  - severity: critical
    area: weights
    description: "All 84 attention projection bias tensors are non-zero while Qwen2 initializes them to zero (layer-0 k_proj.bias absmean 27.7 / max 354). Degenerate generation: whitespace loops at temp <= 0.7 on all 3 seeds."
    remediation: "Re-merge coder-browser LoRA with correct bias handling (do not write adapter biases into base model's zero-initialized projections), re-convert GGUF, re-benchmark."
  - severity: info
    area: collection
    description: "Model pos 34/39 in collection. Duplicate dataset-type repo Nanthasit/sakthai-coder-browser (pos 37, 0 dl) creates naming collision but not inflation."
    remediation: "Consider renaming the dataset-type repo for clarity."