Downloads · 30 days
0
pra-qwen3-5-27b-mlx-4bit
EInnovator/pra-qwen3-5-27b-mlx-4bit
pra-qwen3-5-27b-mlx-4bit is a machine learning model from EInnovator. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for pra. The card lists the license as apache-2.0.
This repository packages the model-specific Progressive Retrieval Attention (PRA) structural mapping, runtime profiles, optional learned components, compatibility metadata, and measured qualification evidence. It does…
Access
Public
Updated Sep 7, 2026
Repo size
5.2 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pt5.2 MB · 94%
At a glance
- Library
- pra
- License
- apache-2.0
- Access
- Public
- Created
- Sep 7, 2026
- Updated
- Sep 7, 2026
- SHA
- 48d50f2b
Base models
Model card
From the Hugging Face model README
PRA Runtime Bundle for mlx-community/Qwen3.5-27B-4bit · MLX / 4bit
What this PRA Runtime Bundle is
This repository packages the model-specific Progressive Retrieval Attention (PRA) structural mapping, runtime profiles, optional learned components, compatibility metadata, and measured qualification evidence. It does not contain the base-model weights and is not an ordinary LoRA quality fine-tune.
- Base model:
mlx-community/Qwen3.5-27B-4bit - Immutable revision:
45797d2985a12c55e6473686e9ea91b95e959553 - Architecture:
Qwen3_5ForConditionalGeneration - Parameters:
27B - Tokenizer revision:
45797d2985a12c55e6473686e9ea91b95e959553 - Serving precision:
INT4/MLX-4bit - Post-training:
pretrained and post-trained
Recommended configuration
- Engine: mlx
- Recommended PRA mode: Selected Context
- Recommended profile: BALANCED
- Bundle evidence tier: CONTROLLED
- Native Memory status: UNAVAILABLE_HYBRID_STATE
Availability, qualification, and recommendation are separate. A mode may be implemented without being qualified or recommended for this identity.
Precision qualification
Precision evidence is scoped to the exact model conversion, engine, mode, and profile. Qualification does not transfer automatically between BF16, INT8, INT4, or encoding-specific formats.
| Family | Encoding | Serving | Feature extraction | Adaptor parameters | Engine | Mode | Profile | Evidence | Datasets |
|---|---|---|---|---|---|---|---|---|---|
| INT4 | MLX-4bit | INT4 | FP32 | FP32 | mlx | Selected Context | BALANCED | CONTROLLED | qasper, hotpotqa |
Headline results
| Dataset | Condition | Token F1 | Exact match | Evidence recall | Visible tokens | TTFT mean | Decode tok/s | Peak memory |
|---|---|---|---|---|---|---|---|---|
| hotpotqa (n=16) | FULL_NO_PRA | 0.8869 | 0.6875 | 1 | 1669 | 1.531e+04 ms | 14.42 | 17.27 GiB |
| hotpotqa (n=16) | PRA_GENERIC | 0.6019 | 0.4375 | 0.5058 | 375.3 | 3642 ms | 13.61 | 15.72 GiB |
| hotpotqa (n=16) | PRA_LEARNED | 0.3199 | 0.25 | 0.2923 | 374.7 | 3638 ms | 15.57 | 15.72 GiB |
| hotpotqa (n=16) | PRA_ORACLE_CONTROL | 0.8997 | 0.75 | 1 | 375.2 | 3640 ms | 14.6 | 15.72 GiB |
| qasper (n=16) | FULL_NO_PRA | 0.5321 | 0.5 | 1 | 3648 | 3.294e+04 ms | 5.695 | 17.27 GiB |
| qasper (n=16) | PRA_GENERIC | 0.3854 | 0.375 | 0.3439 | 772.9 | 7095 ms | 2.922 | 15.73 GiB |
| qasper (n=16) | PRA_LEARNED | 0.4593 | 0.4375 | 0.5531 | 773 | 7079 ms | 3.866 | 15.73 GiB |
| qasper (n=16) | PRA_ORACLE_CONTROL | 0.4697 | 0.4375 | 0.9581 | 772.8 | 7042 ms | 5.841 | 15.73 GiB |
These rows compare full visible context with selected visible context; they are not Native Memory measurements. The oracle is an evidence-availability control, not a deployable selector.
Evidence by engine, mode, and profile
Each row identifies the exact runtime surface for which metrics are available. MEASURED counts scalar metrics with real observations; missing profile/mode combinations are not inferred from another row.
| Engine | Mode | Profile | No PRA | Mode / no adaptor | Same mode / bundle | Measured metric groups |
|---|---|---|---|---|---|---|
| mlx | Selected Context | QUALITY | CALIBRATION_PENDING | Selected Context: CALIBRATION_PENDING | Selected Context + Bundle: CALIBRATION_PENDING | CALIBRATION_PENDING |
| mlx | Selected Context | BALANCED | NEEDS_RUN | Selected Context: NEEDS_RUN | Selected Context + Bundle: NOT_APPLICABLE | NEEDS_RUN |
| mlx | Selected Context | ECONOMY | CALIBRATION_PENDING | Selected Context: CALIBRATION_PENDING | Selected Context + Bundle: CALIBRATION_PENDING | CALIBRATION_PENDING |
| mlx | Selected Context | QASPER-LEARNED | NEEDS_RUN | Selected Context: NEEDS_RUN | Selected Context + Bundle: NEEDS_RUN | NEEDS_RUN |
Canonical staged evidence
A complete staged cohort is not packaged for this exact identity.
| Condition | Evidence status |
|---|---|
| No PRA | NEEDS_RUN |
| Selected Context | NEEDS_RUN |
| Selected Context + Bundle | NEEDS_RUN |
Existing selector-frozen Selected Context versus Native Memory measurements remain reported below as transport evidence; they are not silently relabeled as adaptor evidence.
Installation
pip install 'pra-hf[hf-hub,hf-runtime]'
pra doctor
Quickstart
pra inspect mlx-community/Qwen3.5-27B-4bit -e mlx -a EInnovator/pra-qwen3-5-27b-mlx-4bit
pra evaluate mlx-community/Qwen3.5-27B-4bit -e mlx -D qasper -a EInnovator/pra-qwen3-5-27b-mlx-4bit
pra recommend .pra/runs/latest
pra serve mlx-community/Qwen3.5-27B-4bit -e mlx -a EInnovator/pra-qwen3-5-27b-mlx-4bit -p balanced
Profiles
| Profile | Purpose | Routing | Consumer layers | Status | Recommendation |
|---|---|---|---|---|---|
| QUALITY | Candidate maximum-quality profile; held-out calibration is incomplete | generic cosine | 3, 7, 11, 15, 19, 23, 27, 31, 35, 39, 43, 47, 51, 55, 59, 63 | CALIBRATION_PENDING | Not promoted |
| BALANCED | Qualified default preserving the all-eligible consumer geometry | generic cosine | 3, 7, 11, 15, 19, 23, 27, 31, 35, 39, 43, 47, 51, 55, 59, 63 | QUALIFIED | Default |
| ECONOMY | Reduced-consumer candidate; the held-out quality gate has not passed | generic cosine | CALIBRATION_PENDING | CALIBRATION_PENDING | Not promoted |
| QASPER-LEARNED | Research-only learned routing profile qualified only on matched QASPER routing diagnostics | combined-router-d128 | 3, 7, 11, 15, 19, 23, 27, 31, 35, 39, 43, 47, 51, 55, 59, 63 | RESEARCH | Not promoted |
Engine compatibility
| Engine | Selected Context | Native Memory | Native Serving | Recommended today |
|---|---|---|---|---|
| mlx | validated | UNAVAILABLE_HYBRID_STATE | NOT_APPLICABLE | Selected Context with BALANCED |
End-to-end qualification
The bundle packages the selected-context generation cohort summarized in Headline results. Row-level outputs and immutable execution metadata are retained in qualification/end_task_qualification.json.
Native Memory qualification
What remains to be measured: paired Selected Context versus Native Memory quality and serving economics.
Research diagnostics
| Dataset | Router/profile | Metric | Value | Cohort | Evidence |
|---|---|---|---|---|---|
| qasper | balanced | R@20% | 0.3752 | 16 | CONTROLLED |
| qasper | qasper-learned | R@20% | 0.6099 | 16 | CONTROLLED |
| hotpotqa | balanced | R@20% | 0.5183 | 16 | CONTROLLED |
| hotpotqa | qasper-learned | R@20% | 0.3079 | 16 | CONTROLLED |
| combined | balanced | R@20% | 0.4467 | 32 | CONTROLLED |
| combined | qasper-learned | R@20% | 0.4589 | 32 | CONTROLLED |
These are qualification measurements, not guaranteed production performance. Run pra evaluate on your hardware and workload. Engine version, profile, cohort, evidence tier, date, and artifact provenance remain recorded in qualification/ and bundle.yaml.
How to evaluate locally
pra evaluate mlx-community/Qwen3.5-27B-4bit -e mlx -a EInnovator/pra-qwen3-5-27b-mlx-4bit -D qasper -o .pra/runs/qasper
pra recommend .pra/runs/qasper
pra report .pra/runs/qasper --format html
Known limitations
- The learned router improves QASPER but is not uniformly positive on HotpotQA; it is opt-in rather than the bundle default.
- The held-out routing diagnostic contains 16 examples per dataset and supports controlled routing claims only.
- Detached Native Memory is unavailable for this hybrid recurrent/attention topology; the bundle fails closed to Selected Context.
- The qualification identity is the exact 4bit MLX model and revision; it does not transfer automatically to another checkpoint, engine, or quantization.
- The controlled end-task arm contains 16 examples per dataset; routing and generation claims remain dataset-specific.
- Qwen3.5 interleaves 48 Gated DeltaNet layers with 16 full-attention layers. Selected Context and routing are supported, but detached Native Memory is unavailable until recurrent DeltaNet state has an explicit lifecycle and composition contract.
- Base-model and dataset licenses apply separately to the router artifact.
Training/creation
- Datasets:
QASPER and HotpotQA - Train Examples:
48 - Validation Examples:
16 - Held Out Test Examples:
32 - Seeds:
[11, 23, 37, 53, 71] - Selection:
maximum combined validation AUC0-30 - Method:
multi-positive softmax - Parameter Count:
1310720 - Base Revision:
45797d2985a12c55e6473686e9ea91b95e959553
Reproducibility
- PRA commit:
4002da39b64c5f343407ef8de669755ed002e20d - Bundle build commit:
4002da39b64c5f343407ef8de669755ed002e20d - Bundle schema:
2 - PRA package:
0.2.0rc1 - Component fingerprints and file checksums are recorded in
bundle.yaml.
Community/support
- Library
- pra
- License
- apache-2.0
- Created
- Sep 7, 2026
- Updated
- Sep 7, 2026