PRA Runtime Bundle for mlx-community/Qwen3-4B-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/Qwen3-4B-4bit - Immutable revision:
4dcb3d101c2a062e5c1d4bb173588c54ea6c4d25 - Architecture:
Qwen3ForCausalLM - Parameters:
4B - Tokenizer revision:
4dcb3d101c2a062e5c1d4bb173588c54ea6c4d25
Recommended configuration
- Engine: mlx
- Recommended PRA mode: Selected Context
- Recommended profile: BALANCED
- Bundle evidence tier: CONTROLLED
- Native Memory status: AVAILABLE
Availability, qualification, and recommendation are separate. A mode may be implemented without being qualified or recommended for this identity.
Headline results
No paired end-task headline is available for this exact model, revision, quantization, engine, profile, and execution mode. Routing diagnostics below must not be interpreted as application quality.
Installation
pip install 'pra-hf[hf-hub,hf-runtime]'
pra doctor
Quickstart
pra inspect mlx-community/Qwen3-4B-4bit -e mlx -a EInnovator/pra-qwen3-4b-mlx-4bit
pra evaluate mlx-community/Qwen3-4B-4bit -e mlx -D qasper -a EInnovator/pra-qwen3-4b-mlx-4bit
pra recommend .pra/runs/latest
pra serve mlx-community/Qwen3-4B-4bit -e mlx -a EInnovator/pra-qwen3-4b-mlx-4bit -p balanced
Profiles
| Profile | Purpose | Routing | Consumer layers | Status | Recommendation |
|---|---|---|---|---|---|
| QUALITY | Candidate maximum-quality profile; held-out calibration is incomplete | generic cosine | all eligible | CALIBRATION_PENDING | Not promoted |
| BALANCED | Qualified default preserving the all-eligible consumer geometry | generic cosine | all eligible | 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 | all eligible | RESEARCH | Not promoted |
Engine compatibility
| Engine | Selected Context | Native Memory | Native Serving | Recommended today |
|---|---|---|---|---|
| mlx | validated | AVAILABLE | NOT_MEASURED | Selected Context with BALANCED |
| hf | portable | NOT_MEASURED for the full-precision HF counterpart | NOT_MEASURED | Selected Context; exact MLX artifact only |
End-to-end qualification
What remains to be measured: paired end-task quality for this exact bundle identity.
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% | 0.4129 | 16 | CONTROLLED |
| qasper | qasper-learned | R@20% | 0.6399 | 16 | CONTROLLED |
| hotpotqa | balanced | R@20% | 0.3863 | 16 | CONTROLLED |
| hotpotqa | qasper-learned | R@20% | 0.3424 | 16 | CONTROLLED |
| combined | balanced | R@20% | 0.3996 | 32 | CONTROLLED |
| combined | qasper-learned | R@20% | 0.4912 | 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/Qwen3-4B-4bit -e mlx -a EInnovator/pra-qwen3-4b-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.
- Paired natural-QA evidence contains five examples per dataset and supports engine qualification, not production qualification.
- Reduced consumer-layer configurations failed the held-out quality gate; BALANCED therefore retains all eligible layers.
- The qualification identity is the exact 4-bit MLX model and revision; it does not transfer automatically to full-precision Hugging Face weights or another quantization.
- 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:
655360 - Base Revision:
4dcb3d101c2a062e5c1d4bb173588c54ea6c4d25
Reproducibility
- PRA commit:
52ea386f56fc8d81346325619f3523e585abf78d - Bundle build commit:
52ea386f56fc8d81346325619f3523e585abf78d - Bundle schema:
2 - PRA package:
0.2.0rc1 - Component fingerprints and file checksums are recorded in
bundle.yaml.