StellarVoyager/H3-IR-Qwen3.6-27B-LoRA

H3 IR Qwen 27B LoRA Model Card

Status

This repository contains the selected LoRA adapter and its reviewed release metadata. The public base-model ID, immutable revision, and derivative-license terms are verified. The corrected staging configuration passes full tensor-shape checks and sampled exact-merge reconstruction. Complete private repository verification and document the public MS-SWIFT bridge loading procedure before the first public release.

Model Description

The model converts a video-generation request plus ordered image, video, and audio evidence into a MiniMax-H3 Context-IR prompt. Audio waveforms are not passed directly to the IR model. They are first converted into structured timeline text containing transcripts, timestamps, speakers, sound events, and music descriptions.

• Base model: Qwen/Qwen3.6-27B
• Base revision: 6a9e13bd6fc8f0983b9b99948120bc37f49c13e9
• Base-model license: Apache-2.0
• Method: LoRA
• Rank: 16
• Alpha: 32
• Dropout: 0.05
• Learning rate: 0.0001
• Training epochs: 2
• Selected checkpoint step: 478
• Selection key: candidate_audio_text_incremental1200_plus56_lora_lr0p0001_step478
• Adapter size: 233,607,432 bytes
• Adapter SHA-256: 3ec7ad70243c27b7c5442eb7b548550d6aaa242915444edea7d1eb75830d3142

The training runtime used MS-SWIFT 4.1.3 at revision c6875ef6a962e83f01138bb239b5fb4e5e55b37f and the recorded container image digest sha256:e2a893217186f27dcc6899a53c71d1264b55e362f37268404e2244fd9946149c.

The lower validation-loss run was not automatically selected. Candidates first had to pass 153-case strict format and reference-coverage gates, then were ranked by blinded semantic evaluation with deterministic tie-breaks. The locked test set was not used for model selection.

Evaluation

Evaluation	Samples	Strict valid	Full reference coverage	Teacher similarity
Candidate validation	153	153 (100%)	153 (100%)	0.47899
Locked test	100	95 (95%)	95 (95%)	0.48005

The 153-case results are selection-time validation metrics. The locked test is the better estimate of residual format and reference failures and must be shown alongside them. Aggregate source data is in evaluation/.

Limitations

• Five locked-test cases failed the strict format or reference contract.
• The model depends on upstream audio transcription and description quality.
• It does not acquire a native audio waveform encoder through this training.
• Official Context-IR labels are not assumed to be error-free ground truth.
• The training checkpoint's generated adapter configuration does not encode the verified public base-model ID or exact revision. The release staging copy must add both values without altering the hash-recorded source file.
• Stock Transformers 4.57.1 does not recognize the base model's qwen3_5 architecture. Training and export used MS-SWIFT's Megatron mcore_bridge; a public adapter loading/export procedure for that runtime is still needed.
• Downstream H3 video quality depends on the H3 backend, sampling parameters, media inputs, and prompt length in addition to IR quality.

License and Release Gate

Repository code and the verified base model are Apache-2.0. Before publishing the adapter, verify the reviewed staging files in a private Hub repository and document a reproducible MS-SWIFT bridge loading/export procedure. The current staging adapter matches all 496 base target modules by shape. A sample covering all 12 module types reconstructs the historical merged BF16 weights exactly.