data/index.json

{
 "scenarios": [
  {
   "key": "bit_flip",
   "title": "bit flip (silent data corruption)",
   "blurb": "One flipped exponent bit in one gradient element. Adam normalises the update away, so the loss curve usually never reacts at all.",
   "truth": 372,
   "bytes": 609898
  },
  {
   "key": "lr_spike_loud",
   "title": "learning-rate spike (loud)",
   "blurb": "The easy case, included on purpose: the loss spikes, so watching the curve works here too.",
   "truth": 372,
   "bytes": 610213
  },
  {
   "key": "lr_spike_subtle",
   "title": "learning-rate drift (subtle)",
   "blurb": "A sustained 4x learning rate. The loss degrades gradually, so 'when did it start?' is genuinely hard by eye.",
   "truth": 372,
   "bytes": 608000
  },
  {
   "key": "overflow",
   "title": "fp8-range overflow",
   "blurb": "Gradients pushed outside an emulated fp8 range produce Inf, which propagates into the optimizer state.",
   "truth": 372,
   "bytes": 461278
  },
  {
   "key": "data_poison",
   "title": "data poisoning (12% of each batch)",
   "blurb": "A small contamination, inside the loss curve's own noise band.",
   "truth": 372,
   "bytes": 607198
  },
  {
   "key": "clean",
   "title": "healthy run (no fault)",
   "blurb": "A control. A good detector must report nothing here.",
   "truth": null,
   "bytes": 609191
  }
 ],
 "n_metrics": 124,
 "metrics": [
  "loss",
  "grad_norm",
  "update_norm",
  "param_norm",
  "grad_max_abs",
  "update_max_abs",
  "param_max_abs",
  "nonfinite_grad",
  "nonfinite_param",
  "sign_flip_rate",
  "cos_grad_prev",
  "update_param_ratio",
  "grad_mean",
  "grad_var",
  "grad_absmean",
  "grad_norm_tensor_max",
  "grad_norm_tensor_min",
  "grad_rms",
  "loss_delta",
  "update_norm_tensor_max",
  "group:all:mean",
  "group:all:var",
  "group:all:max_abs",
  "group:all:grad_l2",
  "group:all:update_l2",
  "group:all:sign_flip",
  "group:all:nonfinite",
  "group:all:upd_param_ratio",
  "group:embed:mean",
  "group:embed:var",
  "group:embed:max_abs",
  "group:embed:grad_l2",
  "group:embed:update_l2",
  "group:embed:sign_flip",
  "group:embed:nonfinite",
  "group:embed:upd_param_ratio",
  "group:attn_qkv:mean",
  "group:attn_qkv:var",
  "group:attn_qkv:max_abs",
  "group:attn_qkv:grad_l2",
  "group:attn_qkv:update_l2",
  "group:attn_qkv:sign_flip",
  "group:attn_qkv:nonfinite",
  "group:attn_qkv:upd_param_ratio",
  "group:attn_out:mean",
  "group:attn_out:var",
  "group:attn_out:max_abs",
  "group:attn_out:grad_l2",
  "group:attn_out:update_l2",
  "group:attn_out:sign_flip",
  "group:attn_out:nonfinite",
  "group:attn_out:upd_param_ratio",
  "group:mlp_in:mean",
  "group:mlp_in:var",
  "group:mlp_in:max_abs",
  "group:mlp_in:grad_l2",
  "group:mlp_in:update_l2",
  "group:mlp_in:sign_flip",
  "group:mlp_in:nonfinite",
  "group:mlp_in:upd_param_ratio",
  "group:mlp_out:mean",
  "group:mlp_out:var",
  "group:mlp_out:max_abs",
  "group:mlp_out:grad_l2",
  "group:mlp_out:update_l2",
  "group:mlp_out:sign_flip",
  "group:mlp_out:nonfinite",
  "group:mlp_out:upd_param_ratio",
  "group:norm:mean",
  "group:norm:var",
  "group:norm:max_abs",
  "group:norm:grad_l2",
  "group:norm:update_l2",
  "group:norm:sign_flip",
  "group:norm:nonfinite",
  "group:norm:upd_param_ratio",
  "group:head:mean",
  "group:head:var",
  "group:head:max_abs",
  "group:head:grad_l2",
  "group:head:update_l2",
  "group:head:sign_flip",
  "group:head:nonfinite",
  "group:head:upd_param_ratio",
  "hist_grad:0",
  "hist_grad:1",
  "hist_grad:2",
  "hist_grad:3",
  "hist_grad:4",
  "hist_grad:5",
  "hist_grad:6",
  "hist_grad:7",
  "hist_grad:8",
  "hist_grad:9",
  "hist_grad:10",
  "hist_grad:11",
  "hist_grad:12",
  "hist_grad:13",
  "hist_grad:14",
  "hist_grad:15",
  "hist_update:0",
  "hist_update:1",
  "hist_update:2",
  "hist_update:3",
  "hist_update:4",
  "hist_update:5",
  "hist_update:6",
  "hist_update:7",
  "hist_update:8",
  "hist_update:9",
  "hist_update:10",
  "hist_update:11",
  "hist_update:12",
  "hist_update:13",
  "hist_update:14",
  "hist_update:15",
  "qnorm:0",
  "qnorm:1",
  "qnorm:2",
  "qnorm:3",
  "qnorm:4",
  "qnorm:5",
  "qnorm:6",
  "qnorm:7"
 ],
 "bytes_per_step": 452,
 "win": 32,
 "preset": "nano",
 "steps": 600,
 "exclude": [
  "lr",
  "step_wall_ms",
  "batch_id_mean",
  "batch_uniq_frac"
 ]
}