01
Live run monitoring
Point RunScope at a log file and it follows the run as it trains — reading the existing history instantly, then streaming every new step. No more pasting console output to see where things stand.
LLM training loss analyzer
See where your loss curve is going before you get there.
RunScope reads any training log, diagnoses run health, extracts your learning-rate schedule, and projects the rest of the curve — with honest confidence bands and the milestones worth watching. A focused desktop tool for anyone training language models.
epos_1_4 · pretrain
115 entries · pretraining
Train loss3.240
Projected @ end3.22 ±0.10
LR scheduleCosine
Run Diagnostics3 checks
PASS Generalization healthy — val tracking down
INFO ppl is validation perplexity, not exp(train)
WARN No decay visible in log yet — assert schedule
Capabilities
A full read on a run, from one log file.
Point RunScope at a training log and it does the analysis a careful engineer would — then keeps going, into territory eyeballing a curve can't reach.
02
Health diagnostics
Divergence, loss spikes, batch noise, plateaus, overfitting, LR discontinuities, and data-logging gaps — each surfaced with severity and a plain-language explanation of the likely cause.
03
Stall & slowdown alerts
RunScope learns your run's cadence and warns you the moment the stream stalls or steps quietly start taking longer than baseline — an early signal of thermal throttling or an I/O bottleneck.
04
Multi-run tabs
Track several runs at once in one window — each tab fully isolated with its own data, views, and live feed. Compare experiments without juggling multiple app windows.
05
Schedule-aware forecasting
Extracts your warmup and decay shape, projects it forward, and fits the loss curve against cumulative learning rate so the projection bends where your schedule bends.
06
Regime detection
Auto-classifies pretraining vs SFT and tunes every threshold to match. A flat curve that's normal for SFT and alarming for pretraining is read correctly in both cases.
07
Milestones & stopping points
Predicts the step and wall-clock time to reach each loss level, and flags the point of diminishing returns — the defensible place to stop if compute matters.
08
Validation tracking
Projects end-of-run validation loss and the train/val gap, and tells you when your best checkpoint is likely already behind you.
09
Interactive charts & export
Loss, learning rate, perplexity, and throughput — with raw and smoothed traces (EMA, moving average, or median), a forecast band, spike markers, log scale, and zoom. Promote any metric to the main view, and export the main chart, every chart at once, or the raw run data — as PNG or JSON.
10
Written run summaries
One click composes a plain-language briefing on the run — its state, the findings that matter, the forecast, and the single most useful next move — ready to paste into Slack, a lab notebook, or an issue. Not a template: it reads the actual shape of your run.
Forecasting
Projections that show their work.
Most dashboards will happily extrapolate a straight line off the end of your curve. RunScope fits a real family of laws, backtests them on held-out data, and refuses to project when the fit isn't trustworthy.
Model selection
Power-law, exponential, and log-linear families are fit and backtested on the last 30% of the run. The family that actually predicted unseen data wins — and its out-of-sample error sets the confidence band, which widens honestly with extrapolation distance.
Backtested fit
Honest bands
Refuses bad fits
Warmup excluded
Cumulative-LR model
Loss is fit against integrated learning rate, not raw steps — so a cosine decay's late-run acceleration is modeled, not missed.
Annealing bonus
When decay is observed, an extra loss drop from annealing is fit from your data and folded into the projection.
Assert the schedule
If decay hasn't shown up in the log yet, tell RunScope the shape and it projects the reduced-LR trajectory for you.
Formats
Whatever your trainer prints, RunScope reads.
Paste or drop
Any log in
JSON metrics arrays, HuggingFace trainer_state.json, or raw console output — auto-detected. Any line with a step and a loss works, including nanoGPT and Lightning styles.
Normalized
One clean schema
Loss, val, perplexity, LR, throughput, tokens, and timing are aligned into a single structure. Missing fields are handled gracefully; same-step rows merge.
Analyzed
Insight out
Diagnostics, an interactive chart, and a full forecast — computed locally, on your machine, from whatever fields your log happened to include.
›
step 500/118326 | loss 6.5989 | lr 8.47e-05 | 0.9k tok/s | seen 0.049B | ETA 3512.4h
Interface
Built for reading runs, not decoration.
Every panel maps to a real question you ask about a training run: is it healthy, where is it going, and when should I stop?
Readout strip
Best train3.317 @ 5,000
Best val3.345 @ 5,000
Progress4.9% · 5,750 / 118,326
Regime
Pretraining · high confidence
auto / pretrain / SFT
Run diagnostics & live alerts
◆PASSNo divergence — smoothed loss at run minimum
◆WARNSteps ~40% slower than baseline — possible throttling
◆CRITOverfitting — val rising while train falls
Losstrain + val
LRschedule fit
Perplexitytracked
Throughputtok/s
Forecast panel
Projected final loss with a confidence band, milestone table, diminishing-returns point, and the caveats that apply.
Controls
- Watch a log file live, or paste and analyze a finished run.
- Open multiple runs in tabs; detach a live run to a snapshot.
- Swap any metric into the main chart; smooth by EMA, MA, or median.
- Copy a written summary; export charts or data as PNG or JSON.
Download
Run it on your own curves.
RunScope ships as a lightweight native desktop app for Windows. Watch your training runs live or analyze finished logs — everything runs entirely on your machine, and your logs never leave it.
Version 1.2.0 · Windows desktop · macOS and Linux builds coming soon.