Datasets:

NagaYu
/

flashback-forensics

run_id stringlengths 12 22	fault_kind stringclasses 6 values	fault_description stringclasses 6 values	expected_behaviour stringclasses 6 values	first_bad_step int64 540 900	n_steps int64 1.2k 1.2k	seed int64 0 2	preset stringclasses 1 value	param_count int64 534k 534k	batch int64 8 8	lr float64 0 0	sketch_bytes_per_step int64 452 452	loss_only_prediction int64 -1 901	loss_only_error int64 -1 1	flashback_prediction int64 540 900	flashback_metric stringclasses 7 values	flashback_votes int64 4 86	final_loss float64 1.06 4.25 ⌀	split stringclasses 1 value
lr_spike_loud__s1	lr_spike_loud	learning rate x60 for one step	loud: the loss spikes within a step or two, so the baseline should also find it	720	1,200	1	tiny	533,760	8	0.003	452	721	1	720	update_norm	61	1.062434	train
lr_spike_loud__s2	lr_spike_loud	learning rate x60 for one step	loud: the loss spikes within a step or two, so the baseline should also find it	900	1,200	2	tiny	533,760	8	0.003	452	901	1	900	group:attn_qkv:upd_param_ratio	43	1.126062	train
lr_spike_subtle__s1	lr_spike_subtle	learning rate x4 sustained to the end of the run	slow: the loss degrades gradually, so 'when did it start' is genuinely hard by eye	720	1,200	1	tiny	533,760	8	0.003	452	-1	-1	720	update_param_ratio	25	1.140905	train
lr_spike_subtle__s2	lr_spike_subtle	learning rate x4 sustained to the end of the run	slow: the loss degrades gradually, so 'when did it start' is genuinely hard by eye	900	1,200	2	tiny	533,760	8	0.003	452	-1	-1	900	hist_update:9	24	1.214551	train
bit_flip__s0	bit_flip	one flipped exponent bit in one gradient element (silent data corruption)	silent: Adam normalises the update away, so the loss curve often never reacts	540	1,200	0	tiny	533,760	8	0.003	452	-1	-1	540	grad_norm	15	1.111216	train
bit_flip__s1	bit_flip	one flipped exponent bit in one gradient element (silent data corruption)	silent: Adam normalises the update away, so the loss curve often never reacts	720	1,200	1	tiny	533,760	8	0.003	452	-1	-1	720	grad_norm	15	1.059853	train
lowprec_overflow__s0	lowprec_overflow	fp8-range overflow in the gradients -> Inf/NaN	abrupt: non-finite counters fire on the exact step	540	1,200	0	tiny	533,760	8	0.003	452	541	1	540	grad_norm	78	null	train
lowprec_overflow__s2	lowprec_overflow	fp8-range overflow in the gradients -> Inf/NaN	abrupt: non-finite counters fire on the exact step	900	1,200	2	tiny	533,760	8	0.003	452	901	1	900	grad_norm	86	null	train
data_poison_heavy__s0	data_poison_heavy	50% of every batch replaced by unlearnable tokens	the loss moves too, but gradient-direction statistics move first	540	1,200	0	tiny	533,760	8	0.003	452	540	0	540	loss	60	4.245008	train
data_poison_heavy__s1	data_poison_heavy	50% of every batch replaced by unlearnable tokens	the loss moves too, but gradient-direction statistics move first	720	1,200	1	tiny	533,760	8	0.003	452	720	0	720	grad_mean	45	4.180503	train
data_poison_subtle__s0	data_poison_subtle	12% of every batch replaced by unlearnable tokens	hard: a small contamination is inside the loss curve's own noise band	540	1,200	0	tiny	533,760	8	0.003	452	540	0	540	loss	4	1.916396	train
data_poison_subtle__s1	data_poison_subtle	12% of every batch replaced by unlearnable tokens	hard: a small contamination is inside the loss curve's own noise band	720	1,200	1	tiny	533,760	8	0.003	452	720	0	720	grad_mean	4	1.869771	train

Flashback Forensics

Labelled telemetry from training runs that were deliberately broken at a known step. Every row is a few hundred bytes of per-step summary statistics; the label is the step at which the fault was actually injected.

The point of the dataset: to make "how early can you tell a run went wrong?" a measurable question instead of an anecdote.

config	rows	one row is
`steps`	21,600	one training step of one run: 128 sketch metrics + labels
`runs`	18	one run: fault kind, ground-truth first-bad step, model config

Splits are by run: no step of a test run appears in train.

Fields (`steps`)

run_id, step -- identity
fault_kind -- one of bit_flip, data_poison_heavy, data_poison_subtle, lowprec_overflow, lr_spike_loud, lr_spike_subtle
128 metric columns -- the Flashback sketch: per-group gradient and update norms, maxima, variances, sign-flip rates, non-finite counters, log-magnitude histograms, and gradient-norm quantiles. Names follow flashback.sketch.SketchSchema: scalars like grad_norm, group metrics like group:attn_qkv:max_abs, histogram bins like hist_update:15.
is_after_first_bad -- 1 from the injected step onward
steps_to_first_bad -- signed distance to ground truth (negative = before)

Baselines measured while building this dataset

| detector | mean |predicted - true| | never detected | |---|---:|---:| | loss curve alone | 0.5 | 6 / 18 | | Flashback consensus bisect | 0.0 | 0 / 18 |

Usage

from datasets import load_dataset

ds = load_dataset("NagaYu/flashback-forensics", "steps", split="test")
runs = load_dataset("NagaYu/flashback-forensics", "runs", split="test")

To reproduce, or to generate more with different faults:

pip install flashback
python scripts/build_forensics.py --scale medium --push-to-hub <you>/flashback-forensics

What is not here

No weights, no gradients, no training data -- only aggregate statistics. The full state history lives in a Flashback delta store, which stays local.

Generation

Model: tiny (0.53M parameters), 1200 steps, 6 fault scenarios x 3 seeds. Faults: learning-rate spikes (loud and subtle), single-bit gradient corruption, fp8-range overflow, and data poisoning at two contamination levels. Generated by Flashback v0.1.0.