justinbalexander/stq-basin-map
STQ basin map — 30,000 served cells The basin map for three builds of LiquidAI/LFM2-8B-A1B: the F16 baseline and two simulated quantization builds (levels7, stq2_0). Each cell is one generation from a baseline sampling configuration, classified as loop / completion / short. Simulated builds. These are F16 containers whose routed-expert tensors hold the format's reconstructed values — not real quantized models. Decode speed would measure F16. Grid 100 temperatures… See the full description on the dataset page: https://huggingface.co/datasets/justinbalexander/stq-basin-map.
STQ basin map — 30,000 served cells
The basin map for three builds of LiquidAI/LFM2-8B-A1B: the F16 baseline and two simulated quantization builds (levels7, stq2_0). Each cell is one generation from a baseline sampling configuration, classified as loop / completion / short.
Simulated builds. These are F16 containers whose routed-expert tensors hold the format's reconstructed values — not real quantized models. Decode speed would measure F16.
Grid
100 temperatures (0.0 → 1.2) x 25 prompt variants (nsub 0 → 24) x 4 seeds = 10,000 cells per arm, 30,000 total. Produced by a resident llama-server with 8 slots, one HTTP request per cell.
Provenance
Known caveat — read before comparing texts
Cells were produced 8 at a time. Concurrency changes the sampling distribution: the top-5 logprobs of the first generated token shift by ~0.27 (argmax unchanged), and only 10 of 25 canary cells produced identical text alone vs batched.
Cell texts are therefore not bit-reproducible under loaded serving. Cell classifications appear to be — 25/25 in the canary. The boundary-population flip rate is being measured separately at 1 slot. Cite this dataset as "the basin under 8-slot serving" until that number lands.
Contents
basin_map_cells.tar.gz the run directory: cells/ (30,000 x .txt + .json), audit/, manifestsPer-cell .json sidecars carry full provenance: model sha, binary sha, git commit, prompt sha, sampling parameters, wall time.
Archive sha256: 0c17f3606c8de77269c99f4bb6fc7808da8b9f4fa488b0d78a6cab1c1963b7b2