{
 "schema": "model-scorecard/1",
 "source": "naija-eval",
 "generatedAt": "2026-11-01T11:33:01Z",
 "scores": [
  {
   "model": "anthropic/claude-sonnet-5.5",
   "task": "translate",
   "metric": "chrf++",
   "value": 51.7,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "en->ha"
  },
  {
   "model": "anthropic/claude-sonnet-5.5",
   "task": "translate",
   "metric": "chrf++",
   "value": 67.2,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "ha->en"
  },
  {
   "model": "anthropic/claude-sonnet-5.5",
   "task": "translate",
   "metric": "chrf++",
   "value": 40.7,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "en->ig"
  },
  {
   "model": "anthropic/claude-sonnet-5.5",
   "task": "translate",
   "metric": "chrf++",
   "value": 60.4,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "ig->en"
  },
  {
   "model": "anthropic/claude-sonnet-5.5",
   "task": "translate",
   "metric": "chrf++",
   "value": 32.3,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "en->yo"
  },
  {
   "model": "anthropic/claude-sonnet-5.5",
   "task": "translate",
   "metric": "chrf++",
   "value": 57.1,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "yo->en"
  },
  {
   "model": "anthropic/claude-sonnet-5.5",
   "task": "translate",
   "metric": "chrf++",
   "value": 59.5,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "avg"
  },
  {
   "model": "anthropic/claude-sonnet-5.5",
   "task": "translate",
   "metric": "chrf++",
   "value": 50.5,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "avg"
  },
  {
   "model": "anthropic/claude-sonnet-5.5",
   "task": "translate",
   "metric": "chrf++",
   "value": 44.7,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "avg"
  },
  {
   "model": "anthropic/claude-sonnet-5.5",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 51.7,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "en->ha"
  },
  {
   "model": "anthropic/claude-sonnet-5.5",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 67.2,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "ha->en"
  },
  {
   "model": "anthropic/claude-sonnet-5.5",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 43.5,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "en->ig"
  },
  {
   "model": "anthropic/claude-sonnet-5.5",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 60.4,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "ig->en"
  },
  {
   "model": "anthropic/claude-sonnet-5.5",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 42.0,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "en->yo"
  },
  {
   "model": "anthropic/claude-sonnet-5.5",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 57.1,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "yo->en"
  },
  {
   "model": "anthropic/claude-sonnet-5.5",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 59.5,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "avg"
  },
  {
   "model": "anthropic/claude-sonnet-5.5",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 52.0,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "avg"
  },
  {
   "model": "anthropic/claude-sonnet-5.5",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 49.5,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "avg"
  },
  {
   "model": "google/gemini-3.8-flash",
   "task": "translate",
   "metric": "chrf++",
   "value": 51.0,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "en->ha"
  },
  {
   "model": "google/gemini-3.8-flash",
   "task": "translate",
   "metric": "chrf++",
   "value": 63.8,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "ha->en"
  },
  {
   "model": "google/gemini-3.8-flash",
   "task": "translate",
   "metric": "chrf++",
   "value": 43.9,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "en->ig"
  },
  {
   "model": "google/gemini-3.8-flash",
   "task": "translate",
   "metric": "chrf++",
   "value": 62.6,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "ig->en"
  },
  {
   "model": "google/gemini-3.8-flash",
   "task": "translate",
   "metric": "chrf++",
   "value": 29.7,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "en->yo"
  },
  {
   "model": "google/gemini-3.8-flash",
   "task": "translate",
   "metric": "chrf++",
   "value": 53.7,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "yo->en"
  },
  {
   "model": "google/gemini-3.8-flash",
   "task": "translate",
   "metric": "chrf++",
   "value": 57.4,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "avg"
  },
  {
   "model": "google/gemini-3.8-flash",
   "task": "translate",
   "metric": "chrf++",
   "value": 53.2,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "avg"
  },
  {
   "model": "google/gemini-3.8-flash",
   "task": "translate",
   "metric": "chrf++",
   "value": 41.7,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "avg"
  },
  {
   "model": "google/gemini-3.8-flash",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 51.0,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "en->ha"
  },
  {
   "model": "google/gemini-3.8-flash",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 63.8,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "ha->en"
  },
  {
   "model": "google/gemini-3.8-flash",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 46.7,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "en->ig"
  },
  {
   "model": "google/gemini-3.8-flash",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 62.6,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "ig->en"
  },
  {
   "model": "google/gemini-3.8-flash",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 41.4,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "en->yo"
  },
  {
   "model": "google/gemini-3.8-flash",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 53.7,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "yo->en"
  },
  {
   "model": "google/gemini-3.8-flash",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 57.4,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "avg"
  },
  {
   "model": "google/gemini-3.8-flash",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 54.7,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "avg"
  },
  {
   "model": "google/gemini-3.8-flash",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 47.5,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "avg"
  },
  {
   "model": "openai/gpt-6.1-sol",
   "task": "translate",
   "metric": "chrf++",
   "value": 51.1,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "en->ha"
  },
  {
   "model": "openai/gpt-6.1-sol",
   "task": "translate",
   "metric": "chrf++",
   "value": 62.2,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "ha->en"
  },
  {
   "model": "openai/gpt-6.1-sol",
   "task": "translate",
   "metric": "chrf++",
   "value": 42.8,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "en->ig"
  },
  {
   "model": "openai/gpt-6.1-sol",
   "task": "translate",
   "metric": "chrf++",
   "value": 62.0,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "ig->en"
  },
  {
   "model": "openai/gpt-6.1-sol",
   "task": "translate",
   "metric": "chrf++",
   "value": 31.3,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "en->yo"
  },
  {
   "model": "openai/gpt-6.1-sol",
   "task": "translate",
   "metric": "chrf++",
   "value": 53.9,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "yo->en"
  },
  {
   "model": "openai/gpt-6.1-sol",
   "task": "translate",
   "metric": "chrf++",
   "value": 56.7,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "avg"
  },
  {
   "model": "openai/gpt-6.1-sol",
   "task": "translate",
   "metric": "chrf++",
   "value": 52.4,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "avg"
  },
  {
   "model": "openai/gpt-6.1-sol",
   "task": "translate",
   "metric": "chrf++",
   "value": 42.6,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "avg"
  },
  {
   "model": "openai/gpt-6.1-sol",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 51.1,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "en->ha"
  },
  {
   "model": "openai/gpt-6.1-sol",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 62.2,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "ha->en"
  },
  {
   "model": "openai/gpt-6.1-sol",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 45.8,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "en->ig"
  },
  {
   "model": "openai/gpt-6.1-sol",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 62.0,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "ig->en"
  },
  {
   "model": "openai/gpt-6.1-sol",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 41.0,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "en->yo"
  },
  {
   "model": "openai/gpt-6.1-sol",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 53.9,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "yo->en"
  },
  {
   "model": "openai/gpt-6.1-sol",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 56.7,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "avg"
  },
  {
   "model": "openai/gpt-6.1-sol",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 53.9,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "avg"
  },
  {
   "model": "openai/gpt-6.1-sol",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 47.5,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "avg"
  },
  {
   "model": "anthropic/claude-haiku-5.5",
   "task": "translate",
   "metric": "chrf++",
   "value": 50.0,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "en->ha"
  },
  {
   "model": "anthropic/claude-haiku-5.5",
   "task": "translate",
   "metric": "chrf++",
   "value": 57.1,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "ha->en"
  },
  {
   "model": "anthropic/claude-haiku-5.5",
   "task": "translate",
   "metric": "chrf++",
   "value": 37.8,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "en->ig"
  },
  {
   "model": "anthropic/claude-haiku-5.5",
   "task": "translate",
   "metric": "chrf++",
   "value": 44.8,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "ig->en"
  },
  {
   "model": "anthropic/claude-haiku-5.5",
   "task": "translate",
   "metric": "chrf++",
   "value": 29.6,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "en->yo"
  },
  {
   "model": "anthropic/claude-haiku-5.5",
   "task": "translate",
   "metric": "chrf++",
   "value": 46.0,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "yo->en"
  },
  {
   "model": "anthropic/claude-haiku-5.5",
   "task": "translate",
   "metric": "chrf++",
   "value": 53.5,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "avg"
  },
  {
   "model": "anthropic/claude-haiku-5.5",
   "task": "translate",
   "metric": "chrf++",
   "value": 41.3,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "avg"
  },
  {
   "model": "anthropic/claude-haiku-5.5",
   "task": "translate",
   "metric": "chrf++",
   "value": 37.8,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "avg"
  },
  {
   "model": "anthropic/claude-haiku-5.5",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 50.0,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "en->ha"
  },
  {
   "model": "anthropic/claude-haiku-5.5",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 57.1,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "ha->en"
  },
  {
   "model": "anthropic/claude-haiku-5.5",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 40.3,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "en->ig"
  },
  {
   "model": "anthropic/claude-haiku-5.5",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 44.8,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "ig->en"
  },
  {
   "model": "anthropic/claude-haiku-5.5",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 41.2,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "en->yo"
  },
  {
   "model": "anthropic/claude-haiku-5.5",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 46.0,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "yo->en"
  },
  {
   "model": "anthropic/claude-haiku-5.5",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 53.5,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "avg"
  },
  {
   "model": "anthropic/claude-haiku-5.5",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 42.5,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "avg"
  },
  {
   "model": "anthropic/claude-haiku-5.5",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 43.6,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "avg"
  },
  {
   "model": "local/natlas",
   "task": "translate",
   "metric": "chrf++",
   "value": 47.4,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "en->ha"
  },
  {
   "model": "local/natlas",
   "task": "translate",
   "metric": "chrf++",
   "value": 51.2,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "ha->en"
  },
  {
   "model": "local/natlas",
   "task": "translate",
   "metric": "chrf++",
   "value": 37.1,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "en->ig"
  },
  {
   "model": "local/natlas",
   "task": "translate",
   "metric": "chrf++",
   "value": 44.2,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "ig->en"
  },
  {
   "model": "local/natlas",
   "task": "translate",
   "metric": "chrf++",
   "value": 24.5,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "en->yo"
  },
  {
   "model": "local/natlas",
   "task": "translate",
   "metric": "chrf++",
   "value": 37.6,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "yo->en"
  },
  {
   "model": "local/natlas",
   "task": "translate",
   "metric": "chrf++",
   "value": 49.3,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "avg"
  },
  {
   "model": "local/natlas",
   "task": "translate",
   "metric": "chrf++",
   "value": 40.7,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "avg"
  },
  {
   "model": "local/natlas",
   "task": "translate",
   "metric": "chrf++",
   "value": 31.1,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "avg"
  },
  {
   "model": "local/natlas",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 47.4,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "en->ha"
  },
  {
   "model": "local/natlas",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 51.2,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "ha->en"
  },
  {
   "model": "local/natlas",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 39.7,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "en->ig"
  },
  {
   "model": "local/natlas",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 44.3,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "ig->en"
  },
  {
   "model": "local/natlas",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 35.6,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "en->yo"
  },
  {
   "model": "local/natlas",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 37.6,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "yo->en"
  },
  {
   "model": "local/natlas",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 49.3,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "avg"
  },
  {
   "model": "local/natlas",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 42.0,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "avg"
  },
  {
   "model": "local/natlas",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 36.6,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "avg"
  },
  {
   "model": "local/llama31",
   "task": "translate",
   "metric": "chrf++",
   "value": 22.7,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "en->ha"
  },
  {
   "model": "local/llama31",
   "task": "translate",
   "metric": "chrf++",
   "value": 40.1,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "ha->en"
  },
  {
   "model": "local/llama31",
   "task": "translate",
   "metric": "chrf++",
   "value": 17.0,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "en->ig"
  },
  {
   "model": "local/llama31",
   "task": "translate",
   "metric": "chrf++",
   "value": 33.7,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "ig->en"
  },
  {
   "model": "local/llama31",
   "task": "translate",
   "metric": "chrf++",
   "value": 7.6,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "en->yo"
  },
  {
   "model": "local/llama31",
   "task": "translate",
   "metric": "chrf++",
   "value": 28.3,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "yo->en"
  },
  {
   "model": "local/llama31",
   "task": "translate",
   "metric": "chrf++",
   "value": 31.4,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "avg"
  },
  {
   "model": "local/llama31",
   "task": "translate",
   "metric": "chrf++",
   "value": 25.4,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "avg"
  },
  {
   "model": "local/llama31",
   "task": "translate",
   "metric": "chrf++",
   "value": 17.9,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "avg"
  },
  {
   "model": "local/llama31",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 22.7,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "en->ha"
  },
  {
   "model": "local/llama31",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 40.1,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "ha->en"
  },
  {
   "model": "local/llama31",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 18.5,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "en->ig"
  },
  {
   "model": "local/llama31",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 33.8,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "ig->en"
  },
  {
   "model": "local/llama31",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 11.0,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "en->yo"
  },
  {
   "model": "local/llama31",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 28.4,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "yo->en"
  },
  {
   "model": "local/llama31",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 31.4,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha",
   "slice": "avg"
  },
  {
   "model": "local/llama31",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 26.1,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig",
   "slice": "avg"
  },
  {
   "model": "local/llama31",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 19.7,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo",
   "slice": "avg"
  },
  {
   "model": "mistralai/mistral-small-3.2-24b-instruct",
   "task": "translate",
   "metric": "chrf++",
   "value": 19.9,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-28",
   "language": "ha",
   "slice": "en->ha"
  },
  {
   "model": "mistralai/mistral-small-3.2-24b-instruct",
   "task": "translate",
   "metric": "chrf++",
   "value": 33.7,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-28",
   "language": "ha",
   "slice": "ha->en"
  },
  {
   "model": "mistralai/mistral-small-3.2-24b-instruct",
   "task": "translate",
   "metric": "chrf++",
   "value": 11.2,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-28",
   "language": "ig",
   "slice": "en->ig"
  },
  {
   "model": "mistralai/mistral-small-3.2-24b-instruct",
   "task": "translate",
   "metric": "chrf++",
   "value": 30.1,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-28",
   "language": "ig",
   "slice": "ig->en"
  },
  {
   "model": "mistralai/mistral-small-3.2-24b-instruct",
   "task": "translate",
   "metric": "chrf++",
   "value": 9.9,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-28",
   "language": "yo",
   "slice": "en->yo"
  },
  {
   "model": "mistralai/mistral-small-3.2-24b-instruct",
   "task": "translate",
   "metric": "chrf++",
   "value": 29.1,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-28",
   "language": "yo",
   "slice": "yo->en"
  },
  {
   "model": "mistralai/mistral-small-3.2-24b-instruct",
   "task": "translate",
   "metric": "chrf++",
   "value": 26.8,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-28",
   "language": "ha",
   "slice": "avg"
  },
  {
   "model": "mistralai/mistral-small-3.2-24b-instruct",
   "task": "translate",
   "metric": "chrf++",
   "value": 20.6,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-28",
   "language": "ig",
   "slice": "avg"
  },
  {
   "model": "mistralai/mistral-small-3.2-24b-instruct",
   "task": "translate",
   "metric": "chrf++",
   "value": 19.5,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-28",
   "language": "yo",
   "slice": "avg"
  },
  {
   "model": "mistralai/mistral-small-3.2-24b-instruct",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 19.9,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-28",
   "language": "ha",
   "slice": "en->ha"
  },
  {
   "model": "mistralai/mistral-small-3.2-24b-instruct",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 33.7,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-28",
   "language": "ha",
   "slice": "ha->en"
  },
  {
   "model": "mistralai/mistral-small-3.2-24b-instruct",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 12.6,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-28",
   "language": "ig",
   "slice": "en->ig"
  },
  {
   "model": "mistralai/mistral-small-3.2-24b-instruct",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 30.1,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-28",
   "language": "ig",
   "slice": "ig->en"
  },
  {
   "model": "mistralai/mistral-small-3.2-24b-instruct",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 14.0,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-28",
   "language": "yo",
   "slice": "en->yo"
  },
  {
   "model": "mistralai/mistral-small-3.2-24b-instruct",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 29.1,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-28",
   "language": "yo",
   "slice": "yo->en"
  },
  {
   "model": "mistralai/mistral-small-3.2-24b-instruct",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 26.8,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-28",
   "language": "ha",
   "slice": "avg"
  },
  {
   "model": "mistralai/mistral-small-3.2-24b-instruct",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 21.4,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-28",
   "language": "ig",
   "slice": "avg"
  },
  {
   "model": "mistralai/mistral-small-3.2-24b-instruct",
   "task": "translate",
   "metric": "chrf++_nomarks",
   "value": 21.6,
   "higherIsBetter": true,
   "n": 100,
   "measuredAt": "2026-10-28",
   "language": "yo",
   "slice": "avg"
  },
  {
   "model": "local/natlas",
   "task": "transcribe",
   "metric": "wer",
   "value": 38.5,
   "higherIsBetter": false,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha"
  },
  {
   "model": "local/natlas",
   "task": "transcribe",
   "metric": "wer_nomarks",
   "value": 38.5,
   "higherIsBetter": false,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha"
  },
  {
   "model": "local/natlas",
   "task": "transcribe",
   "metric": "cer",
   "value": 18.4,
   "higherIsBetter": false,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha"
  },
  {
   "model": "local/natlas",
   "task": "transcribe",
   "metric": "cer_nomarks",
   "value": 18.4,
   "higherIsBetter": false,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha"
  },
  {
   "model": "local/whispersmall",
   "task": "transcribe",
   "metric": "wer",
   "value": 92.1,
   "higherIsBetter": false,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha"
  },
  {
   "model": "local/whispersmall",
   "task": "transcribe",
   "metric": "wer_nomarks",
   "value": 92.1,
   "higherIsBetter": false,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha"
  },
  {
   "model": "local/whispersmall",
   "task": "transcribe",
   "metric": "cer",
   "value": 39.6,
   "higherIsBetter": false,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha"
  },
  {
   "model": "local/whispersmall",
   "task": "transcribe",
   "metric": "cer_nomarks",
   "value": 39.6,
   "higherIsBetter": false,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ha"
  },
  {
   "model": "local/natlas",
   "task": "transcribe",
   "metric": "wer",
   "value": 45.4,
   "higherIsBetter": false,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig"
  },
  {
   "model": "local/natlas",
   "task": "transcribe",
   "metric": "wer_nomarks",
   "value": 39.4,
   "higherIsBetter": false,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig"
  },
  {
   "model": "local/natlas",
   "task": "transcribe",
   "metric": "cer",
   "value": 17.9,
   "higherIsBetter": false,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig"
  },
  {
   "model": "local/natlas",
   "task": "transcribe",
   "metric": "cer_nomarks",
   "value": 14.5,
   "higherIsBetter": false,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "ig"
  },
  {
   "model": "local/natlas",
   "task": "transcribe",
   "metric": "wer",
   "value": 64.7,
   "higherIsBetter": false,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo"
  },
  {
   "model": "local/natlas",
   "task": "transcribe",
   "metric": "wer_nomarks",
   "value": 33.5,
   "higherIsBetter": false,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo"
  },
  {
   "model": "local/natlas",
   "task": "transcribe",
   "metric": "cer",
   "value": 24.6,
   "higherIsBetter": false,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo"
  },
  {
   "model": "local/natlas",
   "task": "transcribe",
   "metric": "cer_nomarks",
   "value": 10.2,
   "higherIsBetter": false,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo"
  },
  {
   "model": "local/whispersmall",
   "task": "transcribe",
   "metric": "wer",
   "value": 113.8,
   "higherIsBetter": false,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo"
  },
  {
   "model": "local/whispersmall",
   "task": "transcribe",
   "metric": "wer_nomarks",
   "value": 112.8,
   "higherIsBetter": false,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo"
  },
  {
   "model": "local/whispersmall",
   "task": "transcribe",
   "metric": "cer",
   "value": 84.0,
   "higherIsBetter": false,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo"
  },
  {
   "model": "local/whispersmall",
   "task": "transcribe",
   "metric": "cer_nomarks",
   "value": 76.9,
   "higherIsBetter": false,
   "n": 100,
   "measuredAt": "2026-10-08",
   "language": "yo"
  }
 ]
}