target_model:
  id: Nanthasit/sakthai-context-1.5b-merged-v2
  pipeline_tag: text-generation
  library_name: transformers
  base_model: Qwen/Qwen2.5-1.5B-Instruct
  downloads: 0
  likes: 0
  private: false
  gated: false
  created: '2026-07-30T10:46:30.000Z'
  last_modified: '2026-07-31T04:45:44.000Z'
  model_age_days: 0.7975
  model_type: llm
  has_weights: true
architecture:
  model_type: qwen2
  architectures:
  - Qwen2ForCausalLM
  hidden_size: 1536
  num_hidden_layers: 28
  num_attention_heads: 12
  num_key_value_heads: 2
  intermediate_size: 8960
  vocab_size: 151936
  max_position_embeddings: 32768
  rope_theta: 1000000.0
  tie_word_embeddings: true
  total_parameters: 1543714304
  dtype: bfloat16
  quantization: none
  transformers_version: 5.14.1
  arch_source: config.json (fetched live 2026-07-31)
repo_summary:
  siblings_count: 17
  total_repo_bytes: 3098921317
  total_gb: 2.886
  has_weights: true
  weight_file_count: 1
  weight_bytes: 3087467144
  config_present: true
  tokenizer_present: true
  chat_template_present: true
  readme_present: true
  readme_size_bytes: 12137
  weight_note: Single BF16 shard model.safetensors (2.875 GiB), clean monolith.
benchmarks:
  model_index_count: 1
  metrics_count: 3
  all_verified: false
  pending_metrics: 0
  entries:
  - dataset: Nanthasit/sakthai-bench-v2
    task: text-generation
    metrics:
    - name: Selection Accuracy
      value: 34.9
    - name: Arguments Accuracy
      value: 44.2
    - name: Strict Accuracy
      value: 34.2
    source: .eval_results/sakthai-bench-v2.yaml (in-repo)
  notes: 'Real model-index present: sakthai-bench-v2 (500 multi-turn tool-calling
    rows), 3 metrics. Repo .eval_results/sakthai-bench-v2.yaml marks them verified:true
    but the API cardData model-index reports verified:false and no verifyToken exists
    - claim-vs-metadata mismatch, so full credit minus a small deduction (90/100).
    Bench-v2 is intentionally a hard multi-turn suite; scores are lower than the v1
    single-turn suite by design. No inference re-run possible in cron env (hosted
    router 400, local 1.5B OOMs in constrained env).'
card_quality:
  license: apache-2.0
  base_model_documented: true
  base_model: Qwen/Qwen2.5-1.5B-Instruct
  tags_count: 16
  tags:
  - qwen2.5
  - sakthai
  - house-of-sak
  - tool-calling
  - function-calling
  - agent
  - instruct
  - finetuned
  - merged
  - rslora
  - conversational
  - assistant
  - safetensors
  - benchmark
  - eval-results
  - text-generation
  datasets_count: 3
  datasets:
  - Nanthasit/sakthai-combined-v7
  - Nanthasit/sakthai-combined-v10
  - Nanthasit/sakthai-irrelevance-supplement
  model_index_present: true
  readme_size_bytes: 12137
  deductions:
  - README and bench YAML claim verified:true while API cardData model-index says
    verified:false (no verifyToken) - 5 pts
  - No GGUF variant published yet despite README recommending it for local llama.cpp
    use - 2 pts
  score: 93
health_score:
  overall: 55.4
  components:
    popularity: 0.0
    momentum: 0.0
    benchmarks: 90
    card_quality: 93
    repo_hygiene: 95
  weights:
    popularity: 0.2
    momentum: 0.2
    benchmarks: 0.25
    card_quality: 0.2
    repo_hygiene: 0.15
  formula_note: 0.20*popularity + 0.20*momentum + 0.25*benchmarks + 0.20*card_quality
    + 0.15*repo_hygiene. Day-1 model with 0 downloads -> popularity/momentum 0. Benchmarks
    carry the score via the real bench-v2 model-index entry.
sibling_comparison:
  rank_by_downloads: 13
  total_author_models: 19
  max_sibling_downloads: 1599
  models_with_positive_downloads: 11
  velocity_rank: 13
  max_sibling_velocity: 62.54
  our_velocity: 0.0
eval_type: metadata_cron
eval_note: 'sakthai-context-1.5b-merged-v2 is the v2 flagship of the SakThai tool-calling
  family: Qwen2.5-1.5B-Instruct QLoRA + rsLoRA merge, 1.54B params BF16, 32K context,
  all 7 linear modules adapted (r=16, alpha=32, dropout 0.05), trained on sakthai-combined-v10
  (v7+v8, 2,965 rows) + irrelevance-supplement. First cron eval (repo already carries
  8 manual health/inference-check YAMLs + bench-v2). Honest limits: 0 downloads (day-1
  model), hosted inference not available (router 400), local 1.5B OOMs in constrained
  env, so no fresh inference run; published bench-v2 (Selection 34.9 / Arguments 44.2
  / Strict 34.2) taken from the repo''s own .eval_results YAML. Recommendation: publish
  a GGUF (Q4_K_M) for llama.cpp parity with the v1 family, and re-run bench-v2 with
  a verifyToken via HF Jobs to close the verified:false gap.'
eval_metadata:
  model: Nanthasit/sakthai-context-1.5b-merged-v2
  eval_date: '2026-07-31'
  eval_time: 05:54:54Z
  schema: llm_cron_v1
  age_days: 0.7975
  days_since_last_update: 0.048
  download_velocity: 0.0
  cron_run: 17