target_model:
  id: Nanthasit/sakthai-context-0.5b-tools
  pipeline_tag: text-generation
  library_name: transformers
  base_model: Qwen/Qwen2.5-0.5B-Instruct
  downloads: 94
  likes: 0
  private: false
  gated: false
  created: 2026-07-04T22:27:20.000Z
  last_modified: 2026-07-31T04:47:54.000Z
  model_age_days: 26.28
  model_type: merged full SFT (494M params, bfloat16, ~942 MB safetensors)
  has_weights: true

architecture:
  model_type: qwen2
  architectures: ["Qwen2ForCausalLM"]
  hidden_size: 896
  num_hidden_layers: 24
  num_attention_heads: 14
  num_key_value_heads: 2
  intermediate_size: 4864
  vocab_size: 151936
  max_position_embeddings: 32768
  total_parameters: 494032768
  dtype: bfloat16
  rope_theta: 1000000
  rope_type: default
  tie_word_embeddings: true
  attention: full (24/24 full_attention layers; sliding_window null — no SWA)
  transformers_version: 5.14.1

repo_summary:
  siblings_count: 9
  total_repo_bytes: 999542790
  total_gb: 0.931
  has_weights: true
  weight_file_count: 1
  weight_files: ["model.safetensors (988,097,824 bytes, ~942 MB BF16)"]
  config_present: true
  tokenizer_present: true
  chat_template_present: true
  readme_present: true
  readme_size_bytes: 16229
  eval_files_present: 1
  eval_files: [".eval_results/benchmark-20260731_015807.yaml"]
  note: >-
    Clean 9-file merged repo — single BF16 safetensors, full tokenizer +
    chat_template.jinja, no junk. GGUF lives in the merged sibling
    (sakthai-context-0.5b-merged), which this README links for CPU users.

benchmarks:
  model_index_count: 1
  metrics_count: 8
  all_verified: false
  pending_metrics: 8
  entries:
    - task: text-generation
      dataset: Nanthasit/sakthai-bench-v2 (500 rows)
      metric: Selection Accuracy
      value: 91.8
      verified: false
    - task: text-generation
      dataset: Nanthasit/sakthai-bench-v2 (500 rows)
      metric: Arguments Accuracy
      value: 45.7
      verified: false
    - task: text-generation
      dataset: Nanthasit/sakthai-bench-v2 (500 rows)
      metric: Strict Accuracy
      value: 45.7
      verified: false
    - task: text-generation
      dataset: Nanthasit/sakthai-bench-v2 (500 rows)
      metric: Held-Out Tool Accuracy
      value: 87.8
      verified: false
    - task: text-generation
      dataset: Nanthasit/sakthai-bench-v2 (500 rows)
      metric: Partial Arguments Credit
      value: 58.1
      verified: false
    - task: text-generation
      dataset: Nanthasit/sakthai-bench-v2 (500 rows)
      metric: Degenerate Outputs
      value: 0
      verified: false
    - task: text-generation
      dataset: sakthai-bench-v2 (200 irrelevance rows)
      metric: Correct Silence (no tools offered)
      value: 100.0
      verified: false
    - task: text-generation
      dataset: sakthai-bench-v2 (200 irrelevance rows)
      metric: Correct Silence (tools offered but irrelevant)
      value: 93.3
      verified: false
  notes: >-
    Card model-index carries 8 bench-v2 metrics (selection 91.8% = 2.3x over
    the v1 40.2% baseline; args 45.7%; held-out 87.8% on web_search +
    get_news_headlines; 0 degenerate). Card claims independently reproducible
    via eval_bench.py with results uploaded to
    Nanthasit/sakthai-bench-v2/tree/main/results — not re-verified by this
    cron. A real single-trial inference check in this repo
    (.eval_results/benchmark-20260731_015807.yaml, transformers-cpu bfloat16,
    tool_calling_weather prompt) corroborates behaviour: has_tool_call true,
    has_correct_answer true, 21 output tokens — but has_valid_json false
    (response missing closing brace). Publish a multi-trial verified pass
    before promoting the 91.8% claim to "verified".

training:
  method: SFT with prompt-masked (completion-only) loss
  base_model: Qwen/Qwen2.5-0.5B-Instruct
  dataset: Nanthasit/sakthai-combined-v7 (2,050 rows after bench exclusion)
  lora_config: "r=16, alpha=32, dropout=0.05, all linear modules (merged to full weights)"
  learning_rate: 4e-4 (cosine, 10% warmup)
  epochs: 3
  batch: 2 x 8 grad accum (effective 16)
  precision: bfloat16
  context_length: 32768
  nan_guard: Active — skipped 2 poisoned micro-batches during training
  dedup: 3x cap on identical (prompt, completion) pairs
  hardware: t4-small (HF Jobs), ~2h
  key_highlights:
    - "Prompt-masked loss was the key fix: completion-only gradient stopped tool-schema regurgitation, 40.2% → 91.8% selection (+51.6pp, 2.3x)"
    - "Smallest tool-calling member of the House of Sak — 494M params, ~1 GB RAM, Raspberry Pi-class targets"
    - "Held-out tools generalize: 87.8% on tools never seen in training"
    - "LoRA adapter merged to full weights (this repo IS the merged artifact)"

card_quality:
  license: apache-2.0
  base_model_documented: true
  base_model: Qwen/Qwen2.5-0.5B-Instruct
  tags_count: 28
  tags:
    - agent
    - conversational
    - ollama
    - transformers
    - small-language-model
    - slm
    - tool-use
    - qwen
    - qwen2.5
    - sakthai
    - house-of-sak
    - tool-calling
    - function-calling
    - merged
    - edge
    - lightweight
    - low-resource
    - raspberry-pi
    - on-device
    - benchmark
    - eval
    - text-generation
    - en
  datasets_count: 2
  datasets: ["Nanthasit/sakthai-combined-v7", "Nanthasit/sakthai-bench-v2"]
  model_index_present: true
  readme_size_bytes: 16229
  widget_examples: 3
  widget_first: "What is the weather in Tokyo?"
  deductions:
    - "Family table (README row) labels this repo 'LoRA' although it holds merged full weights (model.safetensors 942 MB)"
    - "Comparison section still lists several sibling evals as 'being evaluated'"
    - "In-repo single-trial inference check had has_valid_json: false (minor)"
  score: 95

health_score:
  overall: 59.9
  components:
    popularity: 0.9
    momentum: 20
    benchmarks: 88
    card_quality: 95
    repo_hygiene: 98
  weights:
    popularity: 0.20
    momentum: 0.20
    benchmarks: 0.25
    card_quality: 0.20
    repo_hygiene: 0.15

sibling_comparison:
  rank_by_downloads: 10
  total_author_models: 19
  max_sibling_downloads: 1599
  models_with_positive_downloads: 11
  velocity_rank: 11
  max_sibling_velocity: 62.62
  our_velocity: 3.58

eval_type: metadata_cron
eval_note: >-
  Run 14 — first snapshot for sakthai-context-0.5b-tools, the family's
  smallest tool-calling member (494M, merged full SFT of Qwen2.5-0.5B-Instruct)
  and the direct sibling of the #2 model sakthai-context-0.5b-merged. 94
  downloads (rank 10/19), 3.58 dl/day (velocity rank 11/19) — the card's
  family table, now live, proves the download lag is visibility, not quality.
  Strengths: genuinely excellent card (8-metric model-index, 91.8% selection
  = 2.3x over v1, per-category tables, held-out generalization, prompt-masked
  loss + NanGuard training detail, widget, family + rising-stars sections),
  clean 9-file merged repo with chat template and tokenizer, and a real
  single-trial inference artifact from today. Weaknesses: popularity component
  raw-count-capped (0.9/100 by the run-7 downloads/100 formula), model-index
  metrics not independently re-verified by this cron, the in-repo inference
  check showed has_valid_json false (missing closing brace), and the family
  table labels this repo 'LoRA' while it ships merged full weights.
  Recommendations: (1) multi-trial verification pass on the 91.8%/45.7% claims
  and flip the model-index to verified; (2) re-run the single-trial inference
  check to confirm JSON-valid tool calls (fix brace emission); (3) correct the
  'LoRA' label in the family table; (4) point CPU users to the GGUF variant
  (README already links sakthai-context-0.5b-merged) and consider bundling a
  Q4 GGUF here for a zero-hop edge path.

eval_metadata:
  model: Nanthasit/sakthai-context-0.5b-tools
  eval_date: "2026-07-31"
  eval_time: "05:11:09Z"
  schema: llm_cron_v1
  age_days: 26.28
  days_since_last_update: 0.016
  download_velocity: 3.58
  cron_run: 14