syedazeez/glassbox-135m-telemetry

glassbox-135m telemetry: every observable signal from a from-scratch reasoning run

The complete telemetry of training syedazeez/glassbox-135m-base from random weights, then SFT and GRPO on arithmetic, plus four fault-injection runs that test the anomaly detectors. It includes a zero-build "control room" viewer so you can browse it locally.

pip install duckdb pyarrow numpy tokenizers
PYTHONPATH=viewer python -m rmo.dash --runs runs --port 8765    # then open http://localhost:8765

Runs

Run	What	Hardware	Notes
pt135m-v1	pretraining 2.40B tokens + arithmetic anneal	1× H100, 101 min	4,579 steps, 0 alerts, 33% MFU
post135m-v1	SFT (600 steps), then 4 GRPO steps	1× H100	SFT collapsed entropy to ~0.01
rl135m-v2	GRPO from SFT, 6–7 digits	1× H100	stopped: 94–100% zero-signal groups
rl135m-v4-hot	GRPO from SFT at temperature 1.6	1× H100	stopped: still ~0–6% mixed groups
rl135m-v3-zero	GRPO from the base model (R1-Zero style), 80 steps	1× H100	23–56% mixed groups, real signal
fault-clean	30M model, no fault	1× A10G	0 alerts in 917 steps (false-positive check)
fault-lrspike	30× LR for one step at 300	1× A10G	loss + grad-norm spike caught at step 301
fault-badbatch	repeated-n-gram batch at 300	1× A10G	loss, grad-norm, throughput detectors fire at 300
fault-nangrad	NaN injected into a gradient at 300	1× A10G	crit alert, update skipped, run recovers
smoke-l4, smoke-h100	3–6 min stack smoke tests	L4, H100

Layout (per run)

manifest.json      config, params, GPU, torch/CUDA versions, git sha, data hashes, $/h
scalars/*.parquet  step, wall, tokens, stage, key, value   (long format, ~1,000 keys)
hists/*.parquet    step, key, edges, counts, min, max, mean, std
texts/*.parquet    canary generations, surprise-ledger documents (highest-loss held-out docs)
sys/*.parquet      NVML: power, clocks, temp, util, memory, PCIe, decoded throttle reasons; host CPU/RAM
rollouts/*.parquet every RL rollout: token ids, text, sampler + trainer + reference logprobs, entropy,
                   reward, verifier result, step-level validity, faithfulness, advantage, inclusion
events.jsonl       phases, checkpoints, curriculum changes, injections, alerts, incidents
incidents/<step>-<reason>/   batch tokens, doc ids, metric tail, alert payload

Key families (scalars.key)

Prefix	Meaning
train/	loss, z-loss, logsumexp, pre-clip grad norm, clip coefficient, skipped updates
anat/	loss anatomy: loss and token share by class (digit, space, punct, word start/cont, special)
act/	per-module activation RMS, absmax, excess kurtosis, zero fraction
attn/{layer}/	per-head max pre-softmax logit and attention entropy (explicit-score probe pass)
grad/ weight/ update/	per-tensor and per-group gradient RMS, weight RMS, update/weight ratio
optim/, muon/	Adam second-moment and denominator quantiles, eps-dominated fraction; Muon update RMS and orthogonality error
gns/	gradient noise scale (McCandlish 2018 two-batch estimator): B_simple, trSigma, G2
spec/	per-matrix σ_max, stable rank, effective (entropy) rank
lens/L{k}/	logit lens per residual layer: CE, top-1, entropy, KL to final
val/	held-out loss and bits-per-byte per domain
perf/ time/ mem/ cost/	tokens/s, MFU, CUDA-event phase timings, telemetry overhead, memory, live $ estimate
sft/, rl/	SFT span NLLs and eval; GRPO reward/verifier, group classes, KL k1/k2/k3, clip fractions, sampler–trainer mismatch, pass@k by digits, faithfulness

Caveats

• Gradient noise scale: the trend (1.6k → 15k tokens as loss fell) matches theory. The absolute level is 30–100× below Kaplan's critical-batch fit and hasn't been independently validated.
• val/arith_val is in-distribution with the arithmetic training text (98% 13-gram overlap by construction). Use the RL/eval digit levels for generalisation.
• RL steps restart at 0 per stage. Join on (run_id, stage, step).