Start here: model status, where to report issues, and index of older threads
#1
by rmems - opened
Start here. This thread is the index for questions and reports about this model repo. Please file bugs and proposals on GitHub (rmems/Spikenaut-SNN issues); code, training and export changes land there first, and this card is synced from GitHub main.
What is published here
- Live bank: exp-025 Dale health-PASS bank, byte-identical to GitHub
6965e12a(promoted in PR #6, GitHub #47). Files:dataset/merged_v2/(snn_model.jsonplus four Q8.8.memimages). - Card text: matches GitHub
mainat392c1958(synced in PR #9). - Training data: rmems/Spikenaut-SNN-Telemetry.
Status of the older Hub threads
| Thread | Status |
|---|---|
| #2 Monotonic hidden weight artifact | Filed against the pre-#47 bank. The linear ramp is no longer in dataset/merged_v2/; see the card's Known limitations and GitHub #2. |
| #3 Inhibitory connections and recurrence | Partly addressed: hidden weights are now mixed-sign with outgoing Dale (neurons 12β15 inhibitory on the readout). Recurrence is still open on GitHub #3. |
| #4 Q8.8 export roundtrip fidelity | Export is now checked and reproduces all four .mem files byte-for-byte, and float-vs-Q8.8 Hamming is published as a measurement. FPGA spike/action/membrane parity is still not measured; GitHub #6, #15, #39. |
Naming note
The GPU safety supervisor referred to as thalamic-relay in the card is now rmems/vahtisiru. It is a peer process, not a dependency of this model.
This post replaces an earlier placeholder ("test issue").
rmems changed discussion title from test to Start here: model status, where to report issues, and index of older threads