decider-4b — hipfire local judge
A hipfire-quantized judge for hipfire serve's /v1/systemone endpoint. It scores typed decisions over supplied options, not free-form chat. Recommended local judge for this endpoint.
Artifact and provenance
| Property | Value |
|---|---|
| File | decider-4b.mq4 |
| Quantization | symmetric MQ4G256V2, uncalibrated |
| Bytes | 2,595,461,120 (2.595 GB decimal) |
| SHA256 | e63520937f16adf2686ee4f00a5a346b4339af36958692bcad7d59359b346948 |
| Upstream | Mapika/decider-4b |
| Upstream revision | 99d7fd5bfcc4d11ba37239dafe0ed653803819d3 |
| Base | Qwen/Qwen3.5-4B-Base |
| License | Apache-2.0; see LICENSE and NOTICE |
| Quantizer build | ca0aee9c925b4180651b194b440bcda80f656ac5 |
Modification notice: hipfire converted the upstream weights using --format mq4v2 --mq4v2-symmetric. No AWQ, imatrix, GPTQ, or calibration was used during quantization. This file differs from the original BF16 weights. Upstream attribution: Mapika; base copyright 2026 Alibaba Cloud. Full Apache-2.0 license and attribution ship alongside the artifact.
Local agreement and latency
These are agreement measurements against a reference judge, not ground-truth accuracy. R9700 (gfx1201), build above, 100 items with three typed questions each:
| Metric | Symmetric MQ4 |
|---|---|
| Top-1 agreement, overall | 85.0% |
| Top-1, yes/no / choice / score | 94% / 88% / 73% |
| 3-question p50 / p95, pooled | 53.7 / 67.8 ms |
| 1-question p50 / p95, pooled | 17.3 / 26.3 ms |
| Probability agreement, raw, 2-fold | 82.2% |
| Probability agreement, temperature-adjusted, 2-fold | 82.6% |
| Peak VRAM in the timing run | 3184 MiB |
Method: six warm-ups; two 3-question passes over 100 items per order repeat, mirrored repeat; pooled 3-question timing includes 400 calls and 1-question timing includes 600 calls. LJ_FIXNULL=1, --max-seq 4096 --spec off. Answers were bit-identical between repeats. The first repeat's pass-2 p50 was 54 ms (3q), 17 ms (1q); second repeat 55/17 ms. Quality uncertainty is substantial at 100 items: the paired top-1 difference against asymmetric MQ4 was +1.0 percentage point, bootstrap 95% interval [-2.0, +4.0].
Post-hoc reference-matching temperatures from RESULTS-table (split seed 7, fit A / score B): yes/no (noul) 1.29, choice 0.99, score 0.84. B-half probability agreement was 82.5% raw / 83.4% adjusted; KL 0.214 / 0.205 nats. The 2-fold estimates above fit on the opposite half. These temperatures are evaluation results, not quantizer calibration and not automatically applied by downloading the artifact. They are distinct from upstream's own temperature configuration.
Usage
With a hipfire build containing the decider:4b registry entry:
hipfire pull decider:4b
hipfire serve --model decider:4b 127.0.0.1:11435 --max-seq 4096 --spec off
curl -sS http://127.0.0.1:11435/v1/systemone \
-H 'Content-Type: application/json' \
-d '{"model":"decider:4b","state":"My card was charged twice for one purchase.","questions":{"department":{"type":"choice","instructions":"Which department should handle this?","criteria":{"billing":"Billing and payments","support":"Technical support","sales":"Sales"}}}}'
The response contains the selected option and its probability distribution. Confidence is model-derived and can be wrong; validate on your own decision distribution before relying on it. Latency depends on hardware, context, options, and runtime settings.