Add sanitized training manifest
Browse files- training_manifest.json +118 -0
training_manifest.json
ADDED
|
@@ -0,0 +1,118 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"schema": "br-voice-reasoner-training-manifest.v2",
|
| 3 |
+
"model": {
|
| 4 |
+
"name": "BR-Voice-Reasoner",
|
| 5 |
+
"release_checkpoint": 2318,
|
| 6 |
+
"format": "merged_full_weights",
|
| 7 |
+
"dtype": "bfloat16",
|
| 8 |
+
"weight_shards": 16
|
| 9 |
+
},
|
| 10 |
+
"base_model": {
|
| 11 |
+
"hub_id": "Qwen/Qwen3-Omni-30B-A3B-Thinking",
|
| 12 |
+
"config_sha256": "e6f8c637528461398ef4e11d48b6112ce73db2458a563dfea548127c2cce253f",
|
| 13 |
+
"index_sha256": "13ae0b6fa947acd1ec5e163d493884724999788752b699266eb09b14ba914c90"
|
| 14 |
+
},
|
| 15 |
+
"training": {
|
| 16 |
+
"method": "OPSD fixed-teacher on-policy distillation",
|
| 17 |
+
"student_initialization": "Qwen/Qwen3-Omni-30B-A3B-Thinking",
|
| 18 |
+
"teacher": {
|
| 19 |
+
"checkpoint": "Qwen/Qwen3-Omni-30B-A3B-Thinking",
|
| 20 |
+
"frozen": true,
|
| 21 |
+
"input_view": "transcript or text form of the same user request with multimodal fields removed",
|
| 22 |
+
"target_answer_available": false
|
| 23 |
+
},
|
| 24 |
+
"steps": 2318,
|
| 25 |
+
"epochs": 2,
|
| 26 |
+
"global_batch_size": 16,
|
| 27 |
+
"learning_rate": 5e-6,
|
| 28 |
+
"projector_learning_rate": 1e-5,
|
| 29 |
+
"minimum_learning_rate": 5e-7,
|
| 30 |
+
"learning_rate_schedule": "cosine",
|
| 31 |
+
"warmup_fraction": 0.03,
|
| 32 |
+
"seed": 20260819,
|
| 33 |
+
"loss": {
|
| 34 |
+
"beta": 0.5,
|
| 35 |
+
"lambda": 1.0,
|
| 36 |
+
"sft_alpha": 0.0,
|
| 37 |
+
"temperature": 1.2,
|
| 38 |
+
"teacher_logits_topk": 128
|
| 39 |
+
},
|
| 40 |
+
"rollout": {
|
| 41 |
+
"source": "student on-policy",
|
| 42 |
+
"thinking": true,
|
| 43 |
+
"maximum_completion_tokens": 2048
|
| 44 |
+
},
|
| 45 |
+
"trainable_modules": [
|
| 46 |
+
"thinker.audio_tower.proj1",
|
| 47 |
+
"thinker.audio_tower.proj2"
|
| 48 |
+
],
|
| 49 |
+
"frozen_components": "all other language-model, audio-encoder, and vision-encoder weights",
|
| 50 |
+
"capability_retention_mechanism": "parameter isolation",
|
| 51 |
+
"optimization_scope_note": "Text-only rows do not update the trainable audio projection modules and are not claimed as a capability-retention mechanism."
|
| 52 |
+
},
|
| 53 |
+
"training_population": {
|
| 54 |
+
"rows": 32965,
|
| 55 |
+
"audio_rows": 25454,
|
| 56 |
+
"text_only_rows": 7511,
|
| 57 |
+
"source_order": "fixed V6-ad order",
|
| 58 |
+
"expanded_or_filtered": false,
|
| 59 |
+
"branches": {
|
| 60 |
+
"ifeval": 6675,
|
| 61 |
+
"mmsu": 6844,
|
| 62 |
+
"retention_instruction_following": 615,
|
| 63 |
+
"retention_knowledge_mcq": 1590,
|
| 64 |
+
"retention_open_response": 3696,
|
| 65 |
+
"retention_reasoning": 968,
|
| 66 |
+
"retention_safety_alignment": 642,
|
| 67 |
+
"sdqa": 11935
|
| 68 |
+
},
|
| 69 |
+
"prompt_only_data_sha256": "d32549b279697d72964e167466b297d67f1ac3de84a728d285d92c09bbc034a8",
|
| 70 |
+
"provenance_sha256": "88e89b8905f0f89c5d1fd3b1915fc6e5d58416277f7da48c59a08ff19065145e"
|
| 71 |
+
},
|
| 72 |
+
"merge": {
|
| 73 |
+
"adapter_checkpoint_sha256": "6bb566463abf4d44730779a4cb8231fc39b733a7aba48f4d29971deecb3d95e7",
|
| 74 |
+
"replacement_semantics": "complete_parameter_replacement",
|
| 75 |
+
"compute_dtype": "float32",
|
| 76 |
+
"output_dtype": "preserve_base_tensor_dtype",
|
| 77 |
+
"changed_parameters": [
|
| 78 |
+
"thinker.audio_tower.proj1.bias",
|
| 79 |
+
"thinker.audio_tower.proj1.weight",
|
| 80 |
+
"thinker.audio_tower.proj2.bias",
|
| 81 |
+
"thinker.audio_tower.proj2.weight"
|
| 82 |
+
],
|
| 83 |
+
"lora_delta_parameters": 0
|
| 84 |
+
},
|
| 85 |
+
"evaluation": {
|
| 86 |
+
"date_utc": "2026-09-02",
|
| 87 |
+
"benchmark": "VoiceBench",
|
| 88 |
+
"protocol_file": "evaluation_protocol.json",
|
| 89 |
+
"empty_generation_recovery": {
|
| 90 |
+
"evaluated_samples": 7783,
|
| 91 |
+
"samples_entering_seed_fallback": 5,
|
| 92 |
+
"samples_recovered": 4,
|
| 93 |
+
"samples_exhausted_and_counted_incorrect": 1,
|
| 94 |
+
"selection": "first non-empty result; no ranking or best-of selection"
|
| 95 |
+
},
|
| 96 |
+
"model_judge": {
|
| 97 |
+
"model": "openai/gpt-4o-mini",
|
| 98 |
+
"model_kind": "alias",
|
| 99 |
+
"accessed_utc": "2026-09-02",
|
| 100 |
+
"gateway": "OpenRouter",
|
| 101 |
+
"upstream_provider_only": "OpenAI",
|
| 102 |
+
"judgments_per_sample": 3
|
| 103 |
+
},
|
| 104 |
+
"nine_subset_mean": 90.4613,
|
| 105 |
+
"non_sdqa_eight_subset_mean": 92.4787,
|
| 106 |
+
"sdqa_gpt4o_mini_majority": {
|
| 107 |
+
"score": 74.3219,
|
| 108 |
+
"correct": 411,
|
| 109 |
+
"total": 553,
|
| 110 |
+
"votes": 3
|
| 111 |
+
},
|
| 112 |
+
"mmsu": {
|
| 113 |
+
"score": 85.6539,
|
| 114 |
+
"correct": 2633,
|
| 115 |
+
"total": 3074
|
| 116 |
+
}
|
| 117 |
+
}
|
| 118 |
+
}
|