---
license: mit
tags:
- mnist
- learning-rate-sweep
- from-scratch
---

# MNIST from-scratch learning-rate sweep — results

10 independent runs of a 1.6M-parameter CNN trained **from random init** on
MNIST ([ylecun/mnist](https://huggingface.co/datasets/ylecun/mnist), 60,000
train / 10,000 test), each for **exactly 60 seconds of wall-clock training**
on a T4 (`t4-small` on HF Jobs). SGD momentum 0.9, batch 128, seed 0
(identical init and batch order across runs), ~173 steps/s (~10.4k steps per
run). Training job: `6ac55085404719ba37662a38` (2026-10-06).

| lr | steps | final train loss | test accuracy | weights |
|---|---|---|---|---|
| 1e-05 | 10394 | 0.5141 | 0.8771 | [abidlabs/mnist-lr-1e-05](https://huggingface.co/abidlabs/mnist-lr-1e-05) |
| 3e-05 | 10410 | 0.2578 | 0.9216 | [abidlabs/mnist-lr-3e-05](https://huggingface.co/abidlabs/mnist-lr-3e-05) |
| 0.0001 | 10406 | 0.1462 | 0.9587 | [abidlabs/mnist-lr-0.0001](https://huggingface.co/abidlabs/mnist-lr-0.0001) |
| 0.0003 | 10413 | 0.0465 | 0.9795 | [abidlabs/mnist-lr-0.0003](https://huggingface.co/abidlabs/mnist-lr-0.0003) |
| 0.001 | 10425 | 0.0135 | 0.9852 | [abidlabs/mnist-lr-0.001](https://huggingface.co/abidlabs/mnist-lr-0.001) |
| 0.003 | 10426 | 0.0014 | 0.9879 | [abidlabs/mnist-lr-0.003](https://huggingface.co/abidlabs/mnist-lr-0.003) |
| **0.01** | 10438 | 0.000134 | **0.9916** | [abidlabs/mnist-lr-0.01](https://huggingface.co/abidlabs/mnist-lr-0.01) |
| 0.03 | 10484 | 1.66e-05 | 0.9909 | [abidlabs/mnist-lr-0.03](https://huggingface.co/abidlabs/mnist-lr-0.03) |
| 0.1 | 10563 | 0.000633 | 0.9896 | [abidlabs/mnist-lr-0.1](https://huggingface.co/abidlabs/mnist-lr-0.1) |
| 1.0 | 500 | NaN (diverged) | 0.0980 | [abidlabs/mnist-lr-1](https://huggingface.co/abidlabs/mnist-lr-1) |

**Best test accuracy: 0.9916 at lr=0.01.** Accuracy falls off on both sides:
too-small LRs undertrain in 60s (1e-05 → 0.877), and lr=1.0 diverged to NaN at
step 500 with momentum 0.9 (accuracy collapsed to 0.098 ≈ chance). lr=0.1
trains fast but slightly less stably near the end than 0.01/0.03.

## Files

- `results.csv` — per-run summary (steps, final loss, test accuracy, weights repo)
- `curves.csv` — train loss every 500 steps per run

## Note on the live dashboard

The dashboard ([abidlabs/mnist-lr-sweep-trackio](https://huggingface.co/spaces/abidlabs/mnist-lr-sweep-trackio))
ingested nothing during this run: its `/api/bulk_log` endpoint returned 500
for every flushed batch — with and without a client token — and its bucket
sync 403'd, in both the first sweep (19:29–19:39 UTC) and the clean re-run
(19:48–19:58 UTC). The bucket's backing DB has not been written since
Oct 2, so the Space's ingest appears wedged server-side. The values here
were parsed from the job log; `replay.py` in the session can repopulate the
dashboard runs from these values once the Space is restarted and healthy.