bench/metrics.json

{
  "params": "7,788,288",
  "val_loss": 3.3579,
  "val_ppl": 28.73,
  "probe_hits": "9/50",
  "probe_accuracy": 0.18,
  "gen_speed_tok_per_s": 19.9,
  "hardware": "8-core ARM, no GPU",
  "samples": [
    {"persona": "analyst", "prompt": "Hi, who are you?", "output": "Identify what is being asked and what would need to be true for popular wrong answers; check the baseline facts."},
    {"persona": "analyst", "prompt": "Explain your method for checking a claim.", "output": "Compare claim against evidence: identify the factual assertion; check who verified the claim and how; it requires a source beyond the claim itself. Verdict: true statement. Confidence: MEDIUM."},
    {"persona": "skeptic", "prompt": "Attack this conclusion: 'Three failures in one week with vans nearby is deliberate sabotage.'", "output": "Checklist: (1) identify the factual assertion; (2) compare against known records; (3) note missing context. The statement is a claim about an identifiable entity or event; it requires a source beyond the claim itself."}
  ]
}