| { |
| "schema": "br-voice-reasoner-training-manifest.v2", |
| "model": { |
| "name": "BR-Voice-Reasoner", |
| "release_checkpoint": 2318, |
| "format": "merged_full_weights", |
| "dtype": "bfloat16", |
| "weight_shards": 16 |
| }, |
| "base_model": { |
| "hub_id": "Qwen/Qwen3-Omni-30B-A3B-Thinking", |
| "config_sha256": "e6f8c637528461398ef4e11d48b6112ce73db2458a563dfea548127c2cce253f", |
| "index_sha256": "13ae0b6fa947acd1ec5e163d493884724999788752b699266eb09b14ba914c90" |
| }, |
| "training": { |
| "method": "OPSD fixed-teacher on-policy distillation", |
| "student_initialization": "Qwen/Qwen3-Omni-30B-A3B-Thinking", |
| "teacher": { |
| "checkpoint": "Qwen/Qwen3-Omni-30B-A3B-Thinking", |
| "frozen": true, |
| "input_view": "transcript or text form of the same user request with multimodal fields removed", |
| "target_answer_available": false |
| }, |
| "steps": 2318, |
| "epochs": 2, |
| "global_batch_size": 16, |
| "learning_rate": 5e-6, |
| "projector_learning_rate": 1e-5, |
| "minimum_learning_rate": 5e-7, |
| "learning_rate_schedule": "cosine", |
| "warmup_fraction": 0.03, |
| "seed": 20260819, |
| "loss": { |
| "beta": 0.5, |
| "lambda": 1.0, |
| "sft_alpha": 0.0, |
| "temperature": 1.2, |
| "teacher_logits_topk": 128 |
| }, |
| "rollout": { |
| "source": "student on-policy", |
| "thinking": true, |
| "maximum_completion_tokens": 2048 |
| }, |
| "trainable_modules": [ |
| "thinker.audio_tower.proj1", |
| "thinker.audio_tower.proj2" |
| ], |
| "frozen_components": "all other language-model, audio-encoder, and vision-encoder weights", |
| "capability_retention_mechanism": "parameter isolation", |
| "optimization_scope_note": "Text-only rows do not update the trainable audio projection modules and are not claimed as a capability-retention mechanism." |
| }, |
| "training_population": { |
| "rows": 32965, |
| "audio_rows": 25454, |
| "text_only_rows": 7511, |
| "source_order": "fixed V6-ad order", |
| "expanded_or_filtered": false, |
| "branches": { |
| "ifeval": 6675, |
| "mmsu": 6844, |
| "retention_instruction_following": 615, |
| "retention_knowledge_mcq": 1590, |
| "retention_open_response": 3696, |
| "retention_reasoning": 968, |
| "retention_safety_alignment": 642, |
| "sdqa": 11935 |
| }, |
| "prompt_only_data_sha256": "d32549b279697d72964e167466b297d67f1ac3de84a728d285d92c09bbc034a8", |
| "provenance_sha256": "88e89b8905f0f89c5d1fd3b1915fc6e5d58416277f7da48c59a08ff19065145e" |
| }, |
| "merge": { |
| "adapter_checkpoint_sha256": "6bb566463abf4d44730779a4cb8231fc39b733a7aba48f4d29971deecb3d95e7", |
| "replacement_semantics": "complete_parameter_replacement", |
| "compute_dtype": "float32", |
| "output_dtype": "preserve_base_tensor_dtype", |
| "changed_parameters": [ |
| "thinker.audio_tower.proj1.bias", |
| "thinker.audio_tower.proj1.weight", |
| "thinker.audio_tower.proj2.bias", |
| "thinker.audio_tower.proj2.weight" |
| ], |
| "lora_delta_parameters": 0 |
| }, |
| "evaluation": { |
| "date_utc": "2026-09-02", |
| "benchmark": "VoiceBench", |
| "protocol_file": "evaluation_protocol.json", |
| "empty_generation_recovery": { |
| "evaluated_samples": 7783, |
| "samples_entering_seed_fallback": 5, |
| "samples_recovered": 4, |
| "samples_exhausted_and_counted_incorrect": 1, |
| "selection": "first non-empty result; no ranking or best-of selection" |
| }, |
| "model_judge": { |
| "model": "openai/gpt-4o-mini", |
| "model_kind": "alias", |
| "accessed_utc": "2026-09-02", |
| "gateway": "OpenRouter", |
| "upstream_provider_only": "OpenAI", |
| "judgments_per_sample": 3 |
| }, |
| "nine_subset_mean": 90.4613, |
| "non_sdqa_eight_subset_mean": 92.4787, |
| "sdqa_gpt4o_mini_majority": { |
| "score": 74.3219, |
| "correct": 411, |
| "total": 553, |
| "votes": 3 |
| }, |
| "mmsu": { |
| "score": 85.6539, |
| "correct": 2633, |
| "total": 3074 |
| } |
| } |
| } |
|
|
