P21 Qwen3-ASR + SEA-LION P4 randomized checkpoint, step 50,315
This private repository preserves the settled inference checkpoint used as the parent of the longer P4 randomized continuation.
Artifact class
Private inference/evaluation checkpoint snapshot. This is not a promoted model release and is not a standalone Transformers AutoModel repository. It contains the three trained BF16 components needed by the P21 composition code:
audio_tower.safetensors: trained Qwen3-ASR audio towertarget_projector.safetensors: trained 1024→4096→5120 projectordecoder_deep_lora.safetensors: 496-module Deep-rsLoRA state for the frozen SEA-LION decoder
The frozen decoder, tokenizer/native-vision assets, and audio processor are external dependencies pinned in checkpoint_config.json. The 14.24 GB optimizer/RNG state, corpus rows/media, evaluator gold and predictions, private paths, and service state are deliberately excluded.
Canonical identity
- Run:
p21_p4_post_sea_randomized_canary_20260911_01 - Generation:
step_050315_1789130103064776203 - P4/global step:
50,315 / 70,330 - Parent generation:
step_049291_1789014139367389200 - Parent state SHA-256:
9c5ac1acf5a9d59242a7de45df10f260597433d0f22d55ce61c634cc64aa0b31 - Checkpoint state SHA-256:
27d54452f2f9cc62755050e7463d02a38ef0efc74c388f4b58dfa7de94eaf446 - Fresh-process strict reload:
PASS_P21_P4_POST_SEA_RANDOMIZED_CANARY_STRICT_RELOAD
See provenance.json and MANIFEST.sha256 for exact component hashes and parameter counts.
Status and evaluation boundary
This snapshot is private, experimental, and not promotion-authorized. Its 1,024-update benchmark was too small to support a promotion decision and showed weaker Tamil and six-language SEA text results than the accepted post-SEA parent. It was retained as a settled continuation parent only after explicit authorization to obtain a larger training signal.
Private Hub storage is archival and evaluation infrastructure. It does not change training-data admission, redistribution, model-release, or promotion state recorded by the source checkpoint.
Composition
The intended runtime composition is:
Pinned Qwen3-ASR processor + trained audio tower
→ trained 1024→4096→5120 projector
→ pinned frozen Qwen-SEA-LION decoder/native-vision stack
→ exact trained 496-module Deep-rsLoRA state
Use the project’s P21 runner/evaluator to construct the architecture, install the Deep-LoRA modules, and strict-load these files. Loading safetensors successfully proves file integrity, not end-to-end model behavior.
Security and privacy
Do not make this repository public or gated-public without a separate rights review and explicit authorization. Do not load pickle-bearing optimizer state from untrusted sources; no optimizer is included here.