evaluation-results.json

{
  "model": "Bharat 24B by Leo Labs",
  "evaluation_date": "2026-10-12",
  "completed_run_cases": 222,
  "reported_scope": "Math and code results only",
  "reported_unique_cases": 192,
  "math": {
    "dataset": "GSM8K fixed128-case subset",
    "correct_numeric_final_answer": 123,
    "cases": 128,
    "accuracy_percent": 96.09375,
    "checker": "MathVerify0.9.0 numeric/symbolic final-answer match",
    "secondary_analysis": true,
    "registered_strict_format_correct": 111,
    "registered_strict_format_accuracy_percent": 86.71875,
    "token_limit_cases": 6
  },
  "code": {
    "dataset": "Original HumanEval fixed64-case subset",
    "correct": 51,
    "cases": 64,
    "pass_at_1_percent": 79.6875,
    "attempts_per_case": 1,
    "greedy_decoding": true,
    "token_limit_cases": 0
  },
  "inference": "Original NF4 backbone plus saved checkpoint625 adapter; native prompts and greedy decoding",
  "math_code_evaluation_wall_time_cuts": 0,
  "training_overlap_normalized_exact_questions": 0,
  "scope_note": "Fixed subsets, not full benchmark or leaderboard results. Math format-tolerant scoring was selected after an earlier format artifact and declared before the remaining100 responses were generated."
}