Erratum E-2026-10-25-01. The SNN energy claims of the micro-kiki programme (0.032 mJ/token, 91.6x) were withdrawn on 2026-10-25. docs/ERRATA.md of hypneum-lab/micro-kiki (entry E-2026-10-25-01) states that the energy figures came from an analytical estimate: "The artifact itself records "status": "theoretical (no weight conversion yet)"." A partial rate-coded conversion of the 35B base was run later (results/spikingkiki-35b-convert.json, 2026-04-21, 8 543 of 31 070 layers) and was never evaluated; the ERRATA entry has not yet been reconciled with that run. The metadata in this repository are published as unvalidated: no conversion result, accuracy or energy claim is attached to them, and matched_snn_key is null for all 408 modules of chat-fr (the only file inspected).

SpikingKiki-V4 — LIF conversion metadata (35 domains)

Metadata-only repository: the 35 lif_metadata.json files that describe the rate-coded LIF conversion of the micro-kiki V4-SOTA LoRA adapters (Phase D of the SpikingKiki track). One file per domain, ~2.7 MB in total.

Why this repository exists

The former repository clemsail/spikingkiki-v4-adapters (137 GB) shipped, for each of the 35 domains, an adapters.safetensors plus this lif_metadata.json. On 2026-10-25 an audit compared the LFS SHA-256 of every adapters.safetensors with the files of micro-kiki-v4-sota : 35 / 35 were byte-identical to the V4-SOTA LoRA weights. (statement of the 2026-10-25 card; no source file in the repository, unverified on 2026-10-31) The "SNN conversion" therefore lives entirely in the JSON metadata (per-module thresholds, time constants, timesteps, matched SNN keys), not in new weights. The duplicate 137 GB were removed ; the metadata is preserved here. (statement of the 2026-10-25 card; no source file in the repository, unverified on 2026-10-31)

Files and format

Hub revision 9bdd6eb2762222f092a4d517fdc320b9b3f4ab74 (last modified 2026-10-25T09:10:09Z), 37 files, 2 692 564 bytes: <domain>/lif_metadata.json for 35 domains, README.md, .gitattributes. The conversion script is scripts/convert_lora_to_snn.py in the source repository (introduced in commit 8233d16, 2026-04-21, "LoRA-to-SNN adapter prototype").

Contents

<domain>/lif_metadata.json   # 35 domains, e.g. chat-fr, components, cpp, devops, ...

Each file maps module names of Qwen3.6-35B-A3B to {threshold, tau, max_rate, timesteps, matched_snn_key}. Defaults observed : timesteps = 128, tau = 1.0, max_rate = 1.0, thresholds per module (e.g. 0.0078125 on linear_attn.in_proj_* of layer 20). Values come from the conversion script, not from a new training run. Checked on chat-fr (2026-10-31): an entry maps a module name such as language_model.model.layers.10.linear_attn.in_proj_a to {matched_snn_key, max_rate, tau, threshold, timesteps}, with matched_snn_key null, max_rate 1.0, tau 1.0, threshold 0.0078125, timesteps 128. Whether any file maps modules to actual SNN weights is unverified.

How to use

  1. Take the LoRA weights from hypneum-lab/micro-kiki-v4-sota (same bytes as the removed adapters.safetensors). The reasoning adapter is not on the Hub (34 of 35 published in micro-kiki-v4-sota); reasoning/lif_metadata.json has no matching weights there.
  2. Apply the per-module LIF parameters from the matching <domain>/lif_metadata.json with scripts/convert_lora_to_snn.py from hypneum-lab/micro-kiki.
  3. The rate-coded base clemsail/spikingkiki-35b-a3b-v4 is incomplete on the Hub (about 32 % of the per-module .npz files were uploaded before the training host was powered down) ; its card says so.

Status and limitations

  • Research artefact of the SpikingKiki track (LAS rate-coded conversion), never evaluated end to end ; no claim of accuracy or energy saving is made.
  • The conversion host (Mac Studio M3 Ultra) is powered down since 2026-10.
  • Evaluation: none published on the Hub. The source README states that energy per token has not been measured and tracks it as hypothesis NH-81 (README.md, section "SNN energy (not measured)"). docs/ERRATA.md records on 2026-10-28 that a 26.5 M BSK spiking LM on the macM1 GPU rail used as much or more energy per token than its dense twin (R = J_dense/J_spiking = 0.91 (95 % CI 0.83-1.01) in generation, 0.73 in batched scoring; NVML replication on the RTX 4090 pending (docs/ERRATA.md, update 2026-10-28)), withdrawing the energy argument for GPU/M1-class hardware.
  • Provenance : source_adapter fields inside each JSON point to the V4-SOTA adapter paths.

Citation

micro-kiki Zenodo concept DOI : 10.5281/zenodo.19666759. Author : Clément Saillant, Hypneum Lab (L'Électron Rare), Grandris, France.

Provenance

  • Source repository: https://github.com/hypneum-lab/micro-kiki (Apache-2.0 LICENSE), checked at commit a3968d9e1f0b6a6b8cc4cb028d4fef3ced24770c.
  • Zenodo DOI: 10.5281/zenodo.19666759, concept DOI of the source repository; no per-artefact DOI.
  • Parent pre-registration: OSF 10.17605/OSF.IO/Q6JYN (CITATION.cff).
  • Hub revision described by the file listing: 9bdd6eb2762222f092a4d517fdc320b9b3f4ab74.
  • Card revision date: 2026-10-31.

AI assistance (transparency)

This repository is developed by a one-person lab with AI assistants used as tools, never as authors:

  • Claude Code (Anthropic) assists with code, documentation, operations and literature search in interactive sessions driven by the PI.
  • Codex CLI (OpenAI, gpt-6.1-sol / gpt-6-astra) runs independent review passes before a release, a pre-registration or a submission; the briefs, reports and verdict tables are published next to the reviewed material whenever a pass is run.

Every scientific claim, verdict and pre-registration decision is taken by the PI; results are reproduced by the test suite and CI, and review reports are evidence, not endorsements. No model is listed as a co-author of a publication or as a co-author of a commit.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for hypneum-lab/spikingkiki-v4-lif-metadata

Adapter
(1)
this model