Add files using upload-large-folder tool

docs/af3_spatial_design.md

@@ -0,0 +1,118 @@
+# AF3 + Spatial-BEATs 设计说明
+
+本目录下的四份新文件实现了把 Spatial-Qwen 的 Spatial 接入范式迁移到
+NVIDIA Audio Flamingo 3 (AF3) 上。
+
+## 核心结论
+
+| 维度 | Spatial-Qwen | Spatial-AF3 |
+|---|---|---|
+| Base LLM hidden | 3584 | **3584**（相同，projector 无需调） |
+| 基座 | Qwen2.5-Omni Thinker/Talker | AF3 = Whisper Enc + **Qwen2-7B Dec** |
+| Spatial 注入机制 | `inputs_embeds.masked_scatter` | 同 |
+| 占位符 | `<\|spatial\|>` | `<spatial>`（新加，token id=151672） |
+| RoPE 定制 | 需要（M-RoPE 3 段） | **不需要**（AF3 用标准 1D RoPE） |
+| Audio 占位符 | `<\|AUDIO\|>` | `<sound>`（AF3 原生） |
+| LoRA 前缀 | `thinker.model` | `language_model.model` |
+
+## 新增文件
+
+| 路径 | 作用 |
+|---|---|
+| `spatial_qwen/model/modeling_spatial_af3.py` | `AudioFlamingo3SpatialForConditionalGeneration`：AF3 子类，新增 `spatial_beats_encoder` + `spatial_beats_projector`，重写 `forward` 按 `<sound>` → `<spatial>` 两次 `masked_scatter` 注入 |
+| `spatial_qwen/model/processing_spatial_af3.py` | `AudioFlamingo3SpatialProcessor`：AF3 Processor 子类，接受 FOA 4ch，W 通道走 Whisper、全 4 通道送给 BEATs，扩展 `<spatial>` 占位符 |
+| `train_spatial_af3_qa.py` | 三阶段训练入口（projector_only / encoder_lora / beats_lora），重用 `train_spatial_beats_qa.py` 里所有与模型无关的工具 |
+| `scripts/bench_test_generate_af3.py` | test split 生成脚本，predictions.jsonl 与 Qwen 路径完全同格式，可直接喂给 `score_test_predictions.py` |
+| `shell/launch_train_spatial_af3_qa.sh` | 启动脚本，默认环境变量 + 分布式包装 |
+
+## 关键设计决策
+
+### 1. transformers fork 依赖
+AF3 需要 NVIDIA 本地 fork（版本 5.0.0rc1，在 spur env 的 4.52.0 下不兼容）。
+训练脚本在 `main()` 起始就把 fork 路径插到 `sys.path` 最前面（`ensure_af3_on_path`），
+确保后续所有 `import transformers.*` 走 fork 版。
+
+启动脚本也会导出 `PYTHONPATH`，并强烈建议**独立 conda env**（`af3`），见 shell 脚本 comment。
+
+### 2. FOA → Whisper 只喂 W 通道
+AF3 的 Whisper encoder 是单声道的，无法直接吃 4ch FOA。Processor 内部对 FOA 样本取
+W（第一声道）送给 Whisper，同时把全 4 通道保留给 Spatial-BEATs encoder。
+
+### 3. 不子类化 `AudioFlamingo3Config`
+`AutoConfig.from_pretrained` 通过 `model_type` 字符串做 registry 查找；
+子类化会导致 registry 冲突。改用**运行时往 config 对象上 attach 字段**的方式
+（与 Qwen 路径的 `tc.spatial_beats_*` 完全同模式）。
+
+### 4. `<sound>` 与 `<spatial>` 两次独立 masked_scatter
+`forward()` 里先跑 AF3 原生的 `<sound>` 注入（抄自 parent），再跑 `<spatial>` 注入。
+两次操作写入不同的 token 位，互不干扰。
+
+### 5. KV-cache decode 时跳过空间编码器
+`prepare_inputs_for_generation` 只在 `cache_position[0] == 0`（prefill）时把
+`spatial_audio` / `spatial_tokens` 挂上，decode 阶段剥离，避免每解码一个 token
+就重跑一遍 BEATs（那会让 generate 变得无法用）。
+
+### 6. Checkpoint 格式与 Qwen 完全一致
+`trainable_state_dict` + optimizer + scheduler + metrics，`score_test_predictions.py`
+不需要任何改动。
+
+## 使用方法
+
+### 1. 准备环境
+```bash
+# 建议独立 env，避免和 Spatial-Qwen 的 transformers 4.52 冲突
+conda create -n af3 python=3.10 -y
+conda activate af3
+cd /apdcephfs_cq10/.../model/transformers && pip install -e .
+pip install peft torch numpy soundfile tensorboard tqdm
+```
+
+### 2. 训练
+```bash
+conda activate af3
+bash shell/launch_train_spatial_af3_qa.sh
+# 只跑 stage2
+START_STAGE=2 bash shell/launch_train_spatial_af3_qa.sh
+# 也要跑 stage3（BEATs 解冻）
+START_STAGE=3 RUN_STAGE3=1 bash shell/launch_train_spatial_af3_qa.sh
+```
+
+### 3. 评测（test split）
+```bash
+torchrun --nproc_per_node=8 scripts/bench_test_generate_af3.py \
+    --run-dir runs/v13d_easy_llmqa_af3/stage2_encoder_lora \
+    --checkpoint-tags best \
+    --qa-root /apdcephfs_cq10/.../easy_filtered \
+    --split test
+python scripts/score_test_predictions.py \
+    --predictions-jsonl runs/v13d_easy_llmqa_af3/stage2_encoder_lora/bench/test/best/predictions.jsonl \
+    --azimuth-threshold-deg 20 --elevation-threshold-deg 10
+```
+
+## 已知 gotcha / 待验证项
+
+1. **Processor 的 `save_pretrained()`**：`AudioFlamingo3Processor` 可能不支持
+   额外的 spatial 字段持久化；训练时已 try/except。读回时 `AutoProcessor` 不会
+   恢复 `spatial_token` 设置，**评测脚本每次重新构造 processor 是正确做法**。
+
+2. **`resize_token_embeddings` 初始化**：新增的 `<spatial>` embedding 行随机初始化。
+   与 Qwen 路径一致；实际推理时会被 BEATs 投影立即覆盖，所以随机值无影响。
+
+3. **BEATs → AF3 可能需要独立的 projector 初始化**：AF3 embed 分布与
+   Qwen2.5-Omni 不同，从 Spatial-Qwen easy 训好的 projector 热启动可能不稳定。
+   首次训练建议**从头 stage1**，跑 ~1 epoch 看 loss 曲线。
+
+4. **长音频 vs `max_audio_len=600`**：AF3 processor 默认支持最长 600s（20 个 30s
+   窗口），但 Spatial-Qwen 的约定是 20s / 16kHz。我们保持 20s 上限（见
+   `spatial_audio_max_seconds=20.0`），`<sound>` 的 token 数对应 ~375 个 token。
+
+## 冒烟测试
+
+建议第一次跑先做 128 个样本的小规模验证：
+```bash
+torchrun --nproc_per_node=1 train_spatial_af3_qa.py \
+    --max-train-samples 128 --max-valid-samples 32 \
+    --epochs 1 --projector-only \
+    --qa-root /apdcephfs_cq10/.../easy_filtered \
+    --output-dir runs/af3_smoke
+```

docs/baseline_experiments.md

@@ -0,0 +1,257 @@
+# Simple Spatial Token Baseline Experiments
+
+This document defines two lightweight spatial-token baselines for comparing
+against the current SELD233-based spatial encoder in the Qwen spatial QA setup.
+
+## Goal
+
+The current strong model path is:
+
+```text
+FOA audio -> DCASE/SELD233 feature bridge -> SELD233 backbone -> spatial adapter
+          -> 2.5 Hz spatial tokens -> projector -> Qwen LLM
+```
+
+For 20 s audio with task `235`, the expected rate is:
+
+```text
+feature frames: 100 Hz, hop_len=160
+SELD frames: 10 Hz, feature_to_seld_ratio=10
+spatial tokens: 2.5 Hz, downsample_factor=4
+20 s segment: 50 spatial tokens
+```
+
+Every baseline below must match the same spatial token rate and prompt contract:
+
+```text
+<|AUDIO|><|spatial|>
+```
+
+The comparison should answer whether the full SELD233 encoder is actually
+needed, or whether simpler spatial cues already explain most of the QA gains.
+
+## Baseline A: Raw FOA Intensity Vector Tokens
+
+Use only the FOA intensity-vector channels as spatial input.
+
+The existing feature bridge already computes baseline-compatible features:
+
+```text
+[B, 7, T_feat, F]
+channels 0..3: W/X/Y/Z log-mel
+channels 4..6: FOA intensity-vector features
+```
+
+For task `235`, `F=128`. The IV baseline should use channels `4..6` only.
+
+Do not use full frequency averaging as the primary baseline. It is too lossy:
+it collapses `[3, F]` IV structure into only 3 numbers per frame before the
+0.4 s temporal pooling step. Keep it only as a lower-bound sanity check.
+
+Recommended default token construction:
+
+```text
+IV features: [B, 3, T_feat, F]
+frequency pooling: pool F into 8 or 16 mel bands -> [B, 3, T_feat, F_band]
+temporal pooling: average every 40 feature frames -> [B, T_spat, 3 * F_band]
+optional window statistics: concat mean/std/max over each 0.4 s window
+token MLP: 3 * F_band * N_stats -> D_token, where D_token matches the current spatial token dim
+projector: reuse the existing spatial projector into Qwen hidden size
+```
+
+For 20 s clips this gives:
+
+```text
+T_feat=2000
+T_spat=ceil(2000 / 40)=50
+```
+
+This is a valid low-capacity baseline. It gives the LLM explicit FOA direction
+cues without using the trained SELD233 temporal encoder.
+
+Suggested IV variants:
+
+- `iv_3d`: mean over all frequency bins, then temporal pool to 2.5 Hz. Use this
+  only as a very weak lower bound.
+- `iv_band8`: pool the 128 mel bins into 8 bands, then temporal pool to 2.5 Hz.
+  This is the recommended first baseline.
+- `iv_band16`: pool into 16 bands. Use this if `iv_band8` is too weak.
+- `iv_band8_stats`: pool into 8 bands and concatenate mean/std over the 0.4 s
+  temporal window. Use this if source activity changes inside the window matter.
+
+Important controls:
+
+- Normalize IV features with the same frontend normalization when possible.
+- Verify FOA axis convention before interpreting azimuth/elevation results.
+- Keep the IV MLP small, otherwise it stops being a "simple feature" baseline.
+- Report parameter count separately from the main SELD233 spatial branch.
+- Report `iv_3d` separately from band-preserving IV. They answer different
+  questions: `iv_3d` tests whether almost no spatial detail is enough; band IV
+  tests whether simple spatial features are enough.
+
+## Baseline B: CNN Neural IV Tokens
+
+Train a small CNN to convert audio-derived spatial features into token vectors.
+
+Recommended input:
+
+```text
+FOA feature tensor [B, 7, T_feat, F]
+or IV-only tensor [B, 3, T_feat, F]
+```
+
+Recommended architecture:
+
+```text
+2D CNN over time-frequency
+temporal downsampling to 2.5 Hz
+MLP projection to D_token
+existing spatial projector to Qwen hidden size
+```
+
+The CNN should be deliberately small. A reasonable first version is:
+
+```text
+Conv2d(C_in -> 32, kernel=3, padding=1)
+GELU
+Conv2d(32 -> 64, kernel=3, stride=(4, 2), padding=1)
+GELU
+frequency pooling
+temporal pooling/resampling to T_spat
+MLP(64 -> D_token)
+```
+
+This baseline answers a different question from raw IV:
+
+```text
+Can a shallow trainable spatial frontend learn enough for QA without the
+pretrained SELD233 backbone?
+```
+
+To avoid making this an unfair hidden strong encoder, keep the CNN shallow and
+train it only under the same QA supervision used by the other spatial QA runs.
+
+## Required Negative Controls
+
+These controls are needed to interpret the baseline results.
+
+- `audio_only`: base Qwen with no spatial tokens.
+- `zero_spatial`: keep `<|spatial|>` placeholders but feed zero vectors.
+- `shuffled_spatial`: feed valid spatial tokens from another sample in the batch
+  or from another file with the same length.
+- `iv_shuffled`: same as raw IV baseline, but shuffle IV tokens across samples.
+
+If `zero_spatial` or `shuffled_spatial` performs close to the real spatial
+baseline, the model is mostly exploiting prompt/answer priors rather than the
+spatial tokens.
+
+## Optional Upper Bounds
+
+These are not required for the first baseline pass, but they are useful if the
+results remain ambiguous.
+
+- `oracle_angle_bin`: feed label-derived angle-bin IDs as tokens. This estimates
+  whether the LLM can use perfect discrete spatial information.
+- `oracle_active_time`: feed label-derived active-time bins. This estimates
+  the ceiling for the simplified v4 time questions.
+- `direct_probe`: train a small classifier on frozen spatial tokens for
+  azimuth/elevation bin and active-at-time tasks. This separates encoder quality
+  from LLM generation quality.
+
+## Training Matrix
+
+Run all baselines on `qa_pairs_v4` first. It avoids direct continuous angle and
+time regression, which was too brittle for causal LM text generation.
+
+Recommended first-pass matrix:
+
+```text
+audio_only
+zero_spatial
+shuffled_spatial
+iv_tokens
+iv_tokens + shuffled control
+cnn_neural_iv
+SELD233 stage3 model
+```
+
+Use the same QA data, LoRA target modules, learning rate range, and checkpoint
+selection rule where possible. For spatial-token baselines, start with:
+
+```text
+train adapter/token frontend + projector + LLM LoRA
+freeze Qwen base weights
+freeze or omit SELD233 backbone
+```
+
+If the raw IV baseline is competitive with the full SELD233 branch, the current
+model may not be using the SELD233 temporal encoder effectively. If the CNN
+baseline is competitive but raw IV is not, then a lightweight trainable frontend
+may be enough for the QA task. If all simple baselines fail but SELD233 improves,
+the pretrained SELD encoder is adding useful structure.
+
+## Metrics
+
+For `qa_pairs_v4`, report metrics by task and question class.
+
+Core metrics:
+
+```text
+azimuth bin choice accuracy
+elevation bin choice accuracy
+active-at-time yes/no accuracy
+active-at-time balanced accuracy
+time-bin choice accuracy
+distance tolerance accuracy
+count accuracy
+source identification accuracy
+```
+
+Do not rely only on overall exact match. For yes/no and multiple-choice tasks,
+also compare against the class-prior baseline and random baseline:
+
+```text
+yes/no random baseline: 50%
+4-way choice random baseline: 25%
+class-prior baseline: majority label accuracy in the test split
+```
+
+## Implementation Notes
+
+The clean implementation point is to add a pluggable spatial-token source before
+the existing projector path:
+
+```text
+--spatial-token-source seld233 | iv | cnn_iv | zero | shuffled
+```
+
+All token sources should return:
+
+```text
+spatial_tokens: [B, T_spat, D_token]
+spatial_token_lengths: [B]
+```
+
+That keeps the existing tokenizer, placeholder expansion, RoPE modal order, and
+Qwen spatial injection path unchanged.
+
+Precomputing token caches is acceptable and probably preferable for the first
+baseline pass:
+
+```text
+precompute_iv_spatial_tokens.py
+precompute_cnn_iv_spatial_tokens.py
+```
+
+The training script can then load these as `spatial_tokens` directly, bypassing
+the SELD233 backbone while preserving the same LLM-side path.
+
+## Main Risks
+
+- Raw IV tokens may encode direction but not source identity. Source-specific
+  questions still require the audio branch and LLM to select the right event.
+- CNN Neural IV can become a strong encoder if it is too wide or too deep.
+- Choice QA can overestimate spatial reasoning if distractors are weak. Use the
+  v4 distractor design and keep the negative controls.
+- Active-at-time QA must be balanced; otherwise a majority-class yes/no model can
+  look good without using spatial or temporal evidence.

docs/paper_outline.md

@@ -0,0 +1,453 @@
+# Spatial-Qwen: Paper Outline (NeurIPS 2025)
+
+> **Working Title**: Spatial-Qwen: Enhancing Large Audio-Language Models with SELD-Pretrained Spatial Encoders for Spatial Audio Understanding
+>
+> **Core Claim**: Injecting a SELD-pretrained spatial encoder as an independent spatial token modality into Qwen2.5-Omni significantly improves spatial audio QA.
+>
+> **Data Policy**: Primary experiments use real data only (STARSS23). Simulated data excluded from main results due to real-sim domain gap; discussed in limitations.
+
+
+---
+
+## KEY BACKGROUND: Closest Existing Work - SPUR (arXiv 2511.06606)
+
+A very similar paper already exists (UMD + Dolby). MUST cite and differentiate in Related Work:
+
+- **Method**: FOA -> Banded Covariance (SSCV) -> Conv3D -> Transformer -> MLP -> injected into LALM audio encoder branch
+- **Benchmark SPUR-Set**: 6 spatial reasoning skills, 12K QA pairs, primarily synthetic data
+- **Results**: Qwen-2.5-Omni w/ SPUR avg=7.38/10; GPT-4o avg=4.50; BAT avg=2.48
+
+**Key differences from our work**:
+
+| Dimension | SPUR | Spatial-Qwen (ours) |
+|-----------|------|---------------------|
+| Spatial encoder | SSCV covariance + Conv3D (scratch) | SELD-pretrained encoder (task-aligned) |
+| Injection method | Into audio encoder branch (mixed with semantics) | Independent spatial modality (separated) |
+| Training data | Mostly synthetic + some real | All real STARSS23 recordings |
+| QA scale | 12K QA, 6 task types | ~178K QA, 12 task types (larger, finer) |
+
+---
+
+## Existing Experimental Numbers
+
+| System | Real Test F | AngE | SELD Score |
+|--------|------------|------|-----------|
+| DCASE baseline (not finetuned) | ~0.35 | ~24 deg | ~0.44 |
+| Task 235 (real only, finetuned) | 0.3705 | 23.6 deg | 0.4562 |
+
+Spatial-BEATs (ov1_local_spatial, valid set):
+- class_acc=39.85%, azi_mae=23.52 deg, ele_mae=9.65 deg, dist_mae=0.475
+
+---
+
+## 1. Abstract (~250 words)
+
+**P1 - Problem**: Sound is 3D. Real-world auditory scenes contain multiple sources from different directions and distances. Existing LALMs process mono/stereo input, cannot perceive spatial information (azimuth, elevation, distance), and perform near-random on spatial audio understanding tasks.
+
+**P2 - Method**: We propose Spatial-Qwen, injecting a SELD-pretrained spatial encoder as an independent spatial token modality into Qwen2.5-Omni, giving the LALM spatial awareness. We also propose SpatialAudioQA - the first large-scale spatial reasoning benchmark based on real FOA recordings (12 task types, ~178K QA pairs).
+
+**P3 - Results**: (fill in after experiments) Comparison vs audio-only baseline; ablation validates spatial token effectiveness.
+
+**P4 - Significance**: Spatial-Qwen bridges the gap between the SELD community (precise localization) and the LALM community (natural language understanding), opening new directions for spatial audio understanding.
+
+---
+
+## 2. Introduction (~1 page)
+
+### 2.1 Opening: Spatial perception is core to auditory intelligence
+
+- Human binaural system naturally localizes sound sources (azimuth, elevation, distance) for navigation, communication, safety
+- FOA (First-Order Ambisonics) is the standard format for capturing 3D soundfields, used by DCASE, VR/AR, robotics
+- Applications: robot sound source tracking, AR/VR immersive audio navigation, hearing aids, autonomous driving
+
+### 2.2 The gap between two communities
+
+**Left side - SELD community has localization but no reasoning**:
+- SELDNet, Spatial-AST etc. precisely output azimuth/elevation/distance coordinates
+- But output is structured coordinates, cannot answer natural language questions:
+  - "Is there someone approaching from my left?"
+  - "Which sound is closest to me?"
+  - "How many sources are active between 5-10 seconds?"
+
+**Right side - LALM community can reason but is spatially blind**:
+- Qwen2-Audio, SALMONN, LTU etc. are strong at semantic understanding but process mono/stereo, discarding spatial cues
+- Near-random on spatial questions (SPUR paper: GPT-4o only 4.50/10)
+- Existing benchmarks (MMAU, AIR-Bench, AudioBench) contain NO spatial reasoning tasks
+
+### 2.3 Limitations of prior work
+
+- **BAT** (ICML 2024): binaural signal + LLaMA-2, limited by HRTF, poor generalization
+- **SPUR** (arXiv 2511.06606): similar goal but injects into audio encoder branch (mixes with semantics), primarily synthetic data, small scale (12K QA)
+- **SpatialVLM/SpatialBot**: successful in vision domain but no audio equivalent
+
+### 2.4 Our approach
+
+- **Key insight**: SELD-pretrained encoder MHSA hidden states already encode rich temporal spatial information, ideal as spatial token source
+- **Solution**: Dual-path parallel - W channel to original audio encoder; 4ch FOA to SELD encoder -> 2.5 Hz spatial tokens -> injected at spatial token positions in LLM
+- **Design principle**: Spatial and semantic paths completely separate, no interference
+
+### 2.5 Contributions
+
+1. **Spatial-Qwen framework**: First method to inject SELD-pretrained spatial encoder as independent spatial token modality into LALM
+2. **SpatialAudioQA Benchmark**: ~178K QA, 12 task types, all based on real FOA recordings (STARSS23)
+3. **Systematic comparison**: DCASE SELD encoder vs. Spatial-BEATs comprehensive encoder comparison
+4. **Ablation study**: Rigorous validation of spatial token effectiveness (audio-only / zero-spatial / shuffled-spatial), and impact on general audio understanding (MMAU)
+5. **Open source**: Code, benchmark data, pretrained weights
+
+---
+
+## 3. Related Work (~1 page)
+
+### 3.1 Large Audio Language Models (LALMs)
+
+| Paper | Year/Venue | Key Contribution | Relation to Our Work |
+|-------|-----------|-----------------|---------------------|
+| AudioPaLM | 2023 arXiv | PaLM-2 + AudioLM tokens | Pioneer audio LLM |
+| Pengi | NeurIPS 2023 | CLAP + GPT-2 unified generation | audio-text baseline |
+| LTU | ICLR 2024 | OpenAQA 1.9M QA | QA generation reference |
+| SALMONN | ICLR 2024 | Dual encoder (Whisper+BEATs)->Q-Former->LLM | Most related architecture |
+| Qwen2-Audio | 2024 arXiv | 7B multi-task | Our base model predecessor |
+| Qwen2.5-Omni | 2025 arXiv | Full multimodal, TMRoPE | Our direct base model |
+| WavLLM | EMNLP 2024 | Dual encoder + LoRA routing | Related dual-path design |
+| Audio Flamingo 3 | Interspeech 2024 | Few-shot, in-context learning | SPUR paper base model |
+| BAT | ICML 2024 | Binaural + LLaMA-2, spatial QA | Most direct prior work |
+
+Core argument: All above methods lack FOA spatial awareness. Our paper fills this gap.
+
+### 3.2 Audio QA Benchmarks
+
+| Paper | Year/Venue | Coverage | Missing |
+|-------|-----------|---------|-------|
+| MMAU | NeurIPS 2024 | 27 audio reasoning types, 10K questions | No spatial localization tasks |
+| AIR-Bench | ACL 2024 | Audio instruction following, 20 types | No spatial tasks |
+| AudioBench | 2024 | 26 dataset universal eval | No spatial tasks |
+| Dynamic-SUPERB | ICASSP 2024 | 55-task collaborative benchmark | No spatial tasks |
+| SPUR-Set | 2024 arXiv | 6 spatial reasoning skills, 12K QA | Mostly synthetic, small scale |
+
+Core argument: SpatialAudioQA is the first real-FOA-based, large-scale (~178K), multi-task spatial audio reasoning benchmark.
+
+### 3.3 Sound Event Localization and Detection (SELD)
+
+| Paper | Year/Venue | Key Contribution |
+|-------|-----------|----------------|
+| SELDNet | JSTSP 2019 | CRNN jointly SED+DOA, foundational |
+| ACCDOA | ICASSP 2021 | Activity-Coupled Cartesian DOA |
+| Multi-ACCDOA | ICASSP 2022 | ADPIT loss, 3-track multi-source (directly used) |
+| STARSS23 | NeurIPS 2023 D&B | Real-scene FOA + 3D annotations (our dataset) |
+| DCASE 2024 Task 3 | DCASE 2024 | Added distance estimation (our encoder source) |
+| Spatial-AST | ICASSP 2023 | ViT for SELD |
+| ELSA | 2024 | FOA-text contrastive learning |
+
+Core argument: SELD produces precise coordinates but cannot answer natural language. Our work converts SELD encoder hidden states to LLM spatial tokens, bridging the two communities.
+
+### 3.4 Spatial Understanding in Multimodal LLMs
+
+| Paper | Year/Venue | Key Contribution |
+|-------|-----------|----------------|
+| SpatialVLM | CVPR 2024 | Visual spatial QA, most architecturally similar |
+| SpatialBot | 2024 | RGB-D encoder + LLM spatial reasoning |
+| SpatialRGPT | NeurIPS 2024 | Spatial region tokens + VLM depth QA |
+| 3D-VisTA | ICCV 2023 | 3D point cloud + text alignment, spatial QA |
+| SPUR | 2024 arXiv | FOA + covariance + Conv3D -> LALM (closest prior) |
+
+Core argument: Visual spatial token injection proven effective (SpatialVLM). Our work extends this paradigm to acoustic spatial domain.
+
+---
+
+## 4. Method (~2 pages)
+
+### 4.1 System Overview
+
+MUST include architecture figure (Fig. 1)
+
+Architecture: Dual-path FOA input. Path A: W-channel mono -> Qwen Audio Encoder (~25Hz) -> Audio Tokens via AUDIO placeholder. Path B: 4ch FOA -> Spatial Encoder (DCASE/BEATs) -> Spatial Tokens [B,50,D_llm] via spatial placeholder. Both injected into Qwen2.5-Omni Thinker LLM.
+
+Content to write:
+- Formalization: given FOA waveform x in R^{T x 4} + question q, output text answer a
+- Dual-path parallel: Path A preserves original audio encoder; Path B adds independent spatial path
+- Token injection: tokenizer expands spatial placeholder to T_spat tokens, thinker forward uses masked_scatter
+
+### 4.2 SpatialAudioQA Benchmark
+
+#### 4.2.1 Data Source
+
+- Base data: STARSS23 real-scene FOA recordings (Sony + Tampere University)
+- Coverage: home, office, public indoor spaces
+- Annotations: per-frame azimuth/elevation/distance + 29 sound classes
+- Segmentation: 20s sliding window, event-aware sampling, scene-based train/val/test split
+
+#### 4.2.2 QA Generation Strategy (v6c)
+
+12 task types (Table 3):
+
+| Task | Description | Answer Format | Example |
+|------|-------------|--------------|-------|
+| identify_source_by_location | Identify source given direction | 4-way MC | "What sound is to your left front?" |
+| estimate_azimuth_bin | Estimate azimuth bin | 4-way MC | "In which direction is the speech?" |
+| estimate_elevation_bin | Estimate elevation | 4-way MC | "Is the footstep above or below you?" |
+| estimate_distance | Estimate distance | numeric (meters) | "How far is the speaker?" |
+| detect_source_active_at_time | Activity at given time | yes/no | "Is music playing at t=5s?" |
+| count_sources | Number of active sources | integer | "How many sounds active at t=8s?" |
+| detect_motion | Is source moving | yes/no | "Is the footstep source moving?" |
+| classify_event_time_bin | Active time period | 4-way MC | "When is the door knock?" |
+| classify_motion | Describe movement | short text | "Describe the movement of the speech source." |
+| compare_azimuth | Compare two source directions | MC | "Is the phone left or right of the speech?" |
+| compare_elevation | Compare two source elevations | MC | "Which sound is higher: A or B?" |
+| compare_distance | Compare two source distances | MC | "Is music closer or farther than speech?" |
+
+Total: ~178K QA pairs across all task types.
+
+### 4.3 Spatial Encoders
+
+#### 4.3.1 DCASE SELD Encoder (Task 235)
+
+Architecture:
+- Backbone: CNN -> bidirectional GRU -> Multi-Head Self-Attention (MHSA) -> FNN output heads
+- Input: FOA 7-channel features [B, 7, T_feat, 128] (4ch log-mel + 3ch intensity vectors)
+- Pretrained: DCASE 2024 Task 3, Multi-ACCDOA + ADPIT loss, STARSS23 real data
+- Spatial token source: final MHSA hidden states [B, T_seld, 128] (NOT ACCDOA head)
+
+Why MHSA hidden (not detection head):
+- Scene-level, temporally ordered, dense representation
+- Retains full class and spatial information, richer than ACCDOA 3D vectors
+- Better suited as LLM conditioning signal
+
+Spatial Adapter: 10 Hz -> 2.5 Hz grouped mean pooling (factor=4); LayerNorm -> Linear(128->256) -> GELU -> Linear(256->256); output [B, T_spat, 256], max 50 tokens for 20s
+
+Projector: MLP: LayerNorm -> Linear(256->512) -> GELU -> Linear(512->4096)
+
+#### 4.3.2 Spatial-BEATs Encoder (ov1_local_spatial)
+
+Architecture: FOA [B,4,T] -> foa_feat [B,7,T_f,128]; W-channel BEATs patch embedding + 7ch SpatialDeltaPatchAdapter -> BEATs trunk -> frequency_pool -> TemporalResampler (2.5Hz) -> semantic_embeddings [B,T_s,768]; LocalSpatialEncoder -> fused_embeddings = LayerNorm(semantic+local) [B,T_s,768]
+
+LLM Bridge: LayerNorm -> 1-layer TransformerEncoder -> Linear(768->1536) -> GELU -> Linear(1536->4096) -> llm_spatial_tokens [B,T_s,4096]
+
+| Property | DCASE SELD | Spatial-BEATs |
+|----------|-----------|---------------|
+| Backbone | CRNN + MHSA | BEATs Transformer |
+| Pretraining objective | SELD Multi-ACCDOA | DOA/distance/classification |
+| Pretraining data | STARSS23 (real) | Simulated FOA |
+| Token source | MHSA hidden [B,T,128] | fused_embeddings [B,T,768] |
+| Token dimension | 128->256->4096 | 768->1536->4096 |
+| Token rate | 2.5 Hz | 2.5 Hz |
+| Parameters | ~5M | ~90M |
+
+### 4.4 Token Injection Mechanism
+
+- Qwen2_5OmniSpatialProcessor: expands spatial token placeholder to T_spat tokens in prompt
+- Thinker _resolve_spatial_tokens() -> projector -> masked_scatter to replace embeddings
+- RoPE: spatial tokens encoded with TMRoPE in [text, audio, spatial] order
+- Inference: inject only during prefill, not repeated during decode
+
+### 4.5 Training
+
+Three-stage training strategy:
+
+| Stage | Frozen | Trainable | Purpose |
+|-------|--------|-----------|-------|
+| Stage 1: spatial_only | SELD encoder + LLM | Adapter + Projector | Establish mapping |
+| Stage 2: adapter_lora | SELD encoder | Adapter + Projector + LoRA | QA fine-tuning (main experiment) |
+| Stage 3: spatial_lora | nothing | SELD tail + Adapter + Projector + LoRA | Joint optimization |
+
+Training config: SpatialAudioQA v6c train set; LoRA rank=8 alpha=16; AdamW lr=3e-5 cosine+5% warmup; batch=1/GPU grad_accum=8 (effective 64); audio=20s 16kHz FOA 4ch
+
+---
+
+## 5. Evaluation Metrics
+
+### 5.1 Spatial QA Metrics
+
+| Task type | Main metric | Notes |
+|-----------|------------|------|
+| Multiple-choice | Accuracy | vs. random baseline 25% |
+| Yes/No | Balanced Accuracy | prevent class imbalance |
+| Distance estimation | MAE (m) + Tolerance Acc (within 0.5m) | |
+| Counting | Accuracy | exact integer match |
+| Motion detection | F1 | |
+| Overall | Weighted Avg Accuracy | |
+| SRS | Acc minus shuffled_Acc | pure spatial reasoning excluding text priors |
+
+### 5.2 SELD Encoder Quality (appendix or section 4.3)
+
+- F-score (20 deg tolerance), AngE, SELD Error (composite), ER20
+
+### 5.3 MMAU Regression Test
+
+Validate that adding spatial encoder does not hurt general audio understanding. Report MMAU Overall Score.
+
+---
+
+## 6. Results (~1.5 pages)
+
+### 6.1 Main QA Results (Table 1)
+
+| System | Spatial Encoder | Azi Acc | Ele Acc | Dist MAE | Active@T BAcc | Count Acc | Overall | SRS |
+|--------|----------------|---------|---------|----------|--------------|-----------|---------|-----|
+| Qwen2.5-Omni (audio-only) | None | ~25% | ~25% | - | ~50% | - | - | 0 |
+| + Zero Spatial | zeros | ~25% | ~25% | - | ~50% | - | - | ~0 |
+| + Shuffled Spatial | DCASE (shuffled) | ~25% | ~25% | - | ~50% | - | - | ~0 |
+| + DCASE Encoder S2 (Ours) | DCASE Task 235 | XX% | XX% | X.Xm | XX% | XX% | XX% | +XX% |
+| + DCASE Encoder S3 | DCASE Task 235 | XX% | XX% | X.Xm | XX% | XX% | XX% | +XX% |
+| + Spatial-BEATs | Spatial-BEATs | XX% | XX% | X.Xm | XX% | XX% | XX% | +XX% |
+
+(Experimental numbers to be filled in)
+
+Key arguments:
+1. audio-only ~= random -> LLM has no spatial awareness
+2. zero/shuffled ~= random -> improvement comes from spatial token information, not text priors
+3. DCASE encoder significantly improves -> SELD pretraining transfers effectively to QA
+
+### 6.2 Ablation Studies (Table 2)
+
+**Ablation A: Spatial Token Validity (essential)**
+
+| Variant | Overall Acc | SRS |
+|---------|------------|-----|
+| audio_only | ~random | 0 |
+| zero_spatial | ~random | ~0 |
+| shuffled_spatial | ~random | ~0 |
+| iv_band8 (raw IV tokens) | ? | ? |
+| cnn_iv (lightweight CNN) | ? | ? |
+| DCASE Encoder S2 (Ours) | best | +XX% |
+
+**Ablation B: Encoder Comparison**
+
+| Encoder | Params | QA Overall | SELD F | AngE |
+|---------|--------|------------|--------|------|
+| Raw IV tokens | ~0.5M | - | N/A | N/A |
+| CNN-IV | ~2M | - | N/A | N/A |
+| DCASE SELD (frozen, S1) | ~5M | - | 0.37 | 23.6 deg |
+| DCASE SELD (finetune S2) | ~5M | - | - | - |
+| Spatial-BEATs (frozen) | ~90M | - | - | 23.5 deg |
+
+**Ablation C: MMAU Regression** - proves independent spatial path does not hurt general audio capability.
+
+**Ablation D: Speed & Parameters** - DCASE ~5M params adds small overhead; Spatial-BEATs ~90M moderate.
+
+### 6.3 QA Analysis (qualitative)
+
+1. Task difficulty stratification: single-source (azimuth) vs. complex (multi-source comparison)
+2. Error analysis: azimuth confusion during overlap; distance error histogram; accuracy by sound class
+3. Success case examples (2-3 QA examples)
+4. SRS analysis: pure spatial reasoning excluding text priors
+
+---
+
+## 7. Conclusion (~0.5 page)
+
+**P1 - Summary**: Spatial-Qwen is the first framework to inject SELD-pretrained encoder as independent spatial token modality into LALM, significantly outperforming audio-only baseline on real FOA spatial QA without degrading general audio understanding.
+
+**P2 - Benchmark contribution**: SpatialAudioQA provides the largest real-scene spatial audio reasoning benchmark to date, 12 task types ~178K QA.
+
+**P3 - Limitations**: Simulated/real domain gap (log-mel mean shift ~0.4-0.5); evaluation limited to STARSS23 indoor scenes; Spatial-BEATs sim pretraining quality.
+
+**P4 - Future work**: Improve sim data quality; extend to outdoor/dynamic scenes; spatial reasoning in multi-turn dialogue.
+
+---
+
+## 8. Broader Impacts (NeurIPS required section)
+
+Positive impacts: hearing-impaired assistive technology; VR/AR immersive experience; robot sound source navigation
+
+Potential risks: Spatial audio surveillance (privacy concerns); validated only on STARSS23 distribution.
+
+---
+
+## 9. Figure and Table Plan
+
+| # | Type | Content | Location |
+|---|------|---------|--------|
+| Fig. 1 | Architecture | Dual-path FOA -> dual encoder -> LLM injection | Section 4.1 |
+| Fig. 2 | Data | SpatialAudioQA task distribution + example QA | Section 4.2 |
+| Fig. 3 | Ablation | audio-only/zero/shuffled/ours bar chart | Section 6.2 |
+| Fig. 4 | Analysis | Per-task accuracy radar + error distribution | Section 6.3 |
+| Table 1 | Main results | All systems, all QA tasks | Section 6.1 |
+| Table 2 | Ablation | Encoder comparison + token validity | Section 6.2 |
+| Table 3 | Dataset stats | SpatialAudioQA 12 task distribution | Section 4.2 |
+| Table 4 | Encoder comparison | DCASE vs. BEATs architecture comparison | Section 4.3 |
+
+---
+
+## 10. Reference List
+
+### Must cite (by importance)
+
+1. Qwen2.5-Omni Technical Report, Qwen Team, arXiv 2025 - our base model
+2. SPUR, arXiv 2511.06606, UMD+Dolby 2024 - closest prior work, MUST cite and differentiate
+3. STARSS23, Shimada & Politis et al., NeurIPS 2023 D&B, arXiv 2306.09196 - our dataset
+4. Multi-ACCDOA, Shimada et al., ICASSP 2022 - DCASE encoder training paradigm
+5. SELDNet, Adavanne et al., JSTSP 2019 - SELD foundation
+6. ACCDOA, Shimada et al., ICASSP 2021 - encoder output representation
+7. DCASE 2024 Task 3, Politis et al. - encoder pretraining task
+8. BEATs, Chen et al., ICML 2023, arXiv 2212.09199 - Spatial-BEATs backbone
+9. SALMONN, Tang et al., ICLR 2024, arXiv 2310.13289 - dual encoder LALM architecture
+10. BAT, Zheng et al., ICML 2024, arXiv 2402.01591 - binaural spatial audio LLM
+11. MMAU, Sakshi et al., NeurIPS 2024, arXiv 2406.09148 - benchmark comparison
+12. SpatialVLM, Chen et al., CVPR 2024, arXiv 2406.13537 - visual spatial token paradigm
+
+### Strongly recommended
+
+13. Qwen2-Audio, Chu et al., arXiv 2407.10759, 2024
+14. LTU, Gong et al., ICLR 2024, arXiv 2305.10790
+15. Pengi, Deshmukh et al., NeurIPS 2023, arXiv 2305.11834
+16. AIR-Bench, Yang et al., ACL 2024, arXiv 2402.07729
+17. AudioPaLM, Rubenstein et al., arXiv 2306.12925, 2023
+18. SpatialRGPT, Cheng et al., NeurIPS 2024
+19. SpatialBot, Cai et al., arXiv 2024
+20. Audio Flamingo 3, Kong et al., Interspeech 2024, arXiv 2402.01831
+21. LoRA, Hu et al., ICLR 2022, arXiv 2106.09685
+
+### Optional
+
+22. WavLLM, Shao et al., EMNLP 2024, arXiv 2404.15708
+23. Hearing Anything Anywhere, Ma et al., CVPR 2024, arXiv 2406.07532
+24. Spatial-AST, Shao et al., ICASSP 2023, arXiv 2211.01246
+25. 3D-VisTA, Zhu et al., ICCV 2023, arXiv 2308.04352
+26. Dynamic-SUPERB, Huang et al., ICASSP 2024, arXiv 2309.09510
+27. AudioBench, Wang et al., 2024, arXiv 2406.16020
+
+---
+
+## 11. NeurIPS 2025 Format Reminders
+
+- Page limit: body <= 9 pages (including figures/tables), references excluded
+- Format: use official neurips_2025.sty
+- Anonymous: double-blind; no author info, affiliations, acknowledgments
+- References: alphabetical order
+- Must include Broader Impacts section
+- Appendix: hyperparameters, training details, dataset details (no page limit)
+
+---
+
+## 12. Writing Strategy Notes
+
+### Narrative thread
+
+1. Spatial perception is core to auditory intelligence -> but ignored by LALM community
+2. SELD has precise localization, LALM has language understanding -> the two do not connect
+3. Solution: SELD encoder hidden states as spatial tokens -> bridge the two communities
+4. SpatialAudioQA provides rigorous large-scale evaluation framework
+5. Ablation rigorously proves: spatial tokens are working, not prompting bias
+
+### Handling simulated data
+
+- Main experiments use only real data (STARSS23)
+- Appendix A.1 Domain Gap Analysis: simulated-real gap (log-mel mean shift ~0.4-0.5)
+
+### Key experiments to run (priority order)
+
+| Experiment | Priority | Est. time |
+|------------|----------|----------|
+| audio-only Qwen2.5-Omni on SpatialAudioQA | P0 | 1 day |
+| zero-spatial / shuffled-spatial ablation | P0 | 0.5 day |
+| DCASE encoder Stage 2 (adapter_lora) | P0 | 2-3 days |
+| IV tokens baseline | P1 | 1 day |
+| MMAU regression test | P1 | 0.5 day |
+| DCASE encoder Stage 3 (spatial_lora) | P2 | 2 days |
+| Spatial-BEATs Stage B | P2 | 2-3 days |
+| CNN-IV baseline | P3 | 1 day |
+
+---
+
+Document version: v2.0, 2025-04-15
+Sources: DCASE/spur-qwen/BEATs docs, SPUR paper arXiv 2511.06606, literature survey

docs/qa_training_smoke.md

@@ -0,0 +1,83 @@
+# SELD233 QA Dataloader Smoke
+
+## Purpose
+Validate that QA `jsonl` records and 20-second FOA audio can be packed into the
+new `audio + spatial + text` input format used by the SELD233 spatial branch.
+
+The smoke script:
+- loads `train.jsonl` and `valid.jsonl`
+- reads FOA waveforms from `audio_path`
+- builds prompts in the form `<|AUDIO|><|spatial|>\n{prompt}\n{answer}`
+- creates `labels` that supervise answer tokens only
+- prints batch tensor shapes and multimodal placeholder counts
+- optionally loads the full Qwen model and runs one no-grad forward
+
+## Script
+- [train_seld233_qa_smoke.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/spur-qwen-2.5-omni/training-qwen-omni/train_seld233_qa_smoke.py)
+
+## Dataloader-Only Smoke
+
+```bash
+/data/home/schmittzhu/miniconda3/envs/spur/bin/python \
+  /apdcephfs_cq10/share_1603164/user/schmittzhu/code/spur-qwen-2.5-omni/training-qwen-omni/train_seld233_qa_smoke.py \
+  --dataloader-only \
+  --max-train-samples 4 \
+  --max-valid-samples 2 \
+  --batch-size 2
+```
+
+Expected outputs include:
+- `input_ids`
+- `input_features`
+- `spatial_audio`
+- `labels`
+- per-sample `audio_tokens`, `spatial_tokens`, and supervised answer token counts
+
+## Full Forward Smoke
+
+Run this on a GPU machine. CPU loading of the 7B model is possible but very
+slow.
+
+```bash
+/data/home/schmittzhu/miniconda3/envs/spur/bin/python \
+  /apdcephfs_cq10/share_1603164/user/schmittzhu/code/spur-qwen-2.5-omni/training-qwen-omni/train_seld233_qa_smoke.py \
+  --smoke-forward \
+  --max-train-samples 1 \
+  --max-valid-samples 1 \
+  --batch-size 1 \
+  --device cuda:0 \
+  --dtype bfloat16
+```
+
+If the forward succeeds, the script prints:
+- `loss`
+- `logits shape`
+
+## Single-Step Training Smoke
+
+Run this on a GPU machine if you want to verify that one backward pass and one
+optimizer step complete successfully. This smoke test freezes the whole model
+except the newly added spatial adapter and spatial projector.
+
+```bash
+/data/home/schmittzhu/miniconda3/envs/spur/bin/python \
+  /apdcephfs_cq10/share_1603164/user/schmittzhu/code/spur-qwen-2.5-omni/training-qwen-omni/train_seld233_qa_smoke.py \
+  --smoke-backward \
+  --max-train-samples 1 \
+  --max-valid-samples 1 \
+  --batch-size 1 \
+  --device cuda:0 \
+  --dtype bfloat16
+```
+
+If the step succeeds, the script prints:
+- `trainable_parameters`
+- `loss`
+- `gradient_norm_samples`
+
+## Notes
+- This QA loader currently expects 4-channel FOA audio only.
+- It assumes 16 kHz audio and crops any longer clip to the first 20 seconds.
+- The current QA dataset already points to 20-second prepared audio, so the
+  first batches should produce `500` audio tokens and `50` spatial tokens per
+  sample.

docs/qwen3_env.md

@@ -0,0 +1,77 @@
+# Qwen3-Omni-MoE Environment Notes
+
+## Conda env: `spur-qwen3` (cloned from `spur` on 2026-05-10)
+
+```
+python      : 3.10.19
+torch       : 2.10.0+cu128 (CUDA 12.8)
+torchaudio  : 2.10.0
+transformers: 5.0.0.dev0  (editable install from local fork)
+huggingface_hub: 1.14.0
+tokenizers  : 0.22.2
+peft        : 0.18.1
+deepspeed   : 0.19.0
+flash_attn  : NOT installed (use sdpa attn_impl)
+```
+
+## Critical paths
+
+- **transformers fork (live, editable)**:
+  `/apdcephfs_cq10/share_1603164/user/schmittzhu/model/transformers/`
+  Installed via `pip install -e .` in spur-qwen3. Both training & bench code import directly from this.
+- **Snapshot of fork before any modifications**:
+  `/apdcephfs_cq10/share_1603164/user/schmittzhu/model/transformers.snapshot-pre-qwen3/`
+- **Qwen3-Omni-30B-A3B-Instruct weights**:
+  `/apdcephfs_cq12/share_302080740/model/Qwen3-Omni-30B-A3B-Instruct/` (66 GB, 15 safetensors)
+
+## Loading caveats (fork bugs we must work around)
+
+1. **Top-level `Qwen3OmniMoeConfig.from_pretrained(MODEL)` crashes** on
+   `code_predictor_config.use_sliding_window` AttributeError (fork bug). Workaround:
+   parse `config.json` raw and instantiate `Qwen3OmniMoeThinkerConfig(**raw["thinker_config"])`
+   directly. We never need the talker, so this is fine.
+2. **`AutoProcessor.from_pretrained` crashes** on missing video_processor mapping. Workaround:
+   load `AutoTokenizer` + `AutoFeatureExtractor` separately and assemble a
+   `Qwen3OmniMoeProcessor(image_processor=..., video_processor=None, feature_extractor=..., tokenizer=..., chat_template=...)`.
+3. flash-attn not installed → use `attn_implementation="sdpa"` everywhere (already what we do for Qwen2.5).
+
+## Qwen3 Thinker config fingerprint (for sanity check)
+
+```
+hidden_size           = 2048    (Qwen2.5: 4096) ← projector output dim must change
+num_hidden_layers     = 48      (Qwen2.5: 28)
+num_experts           = 128
+num_experts_per_tok   = 8
+mrope_section         = (24, 20, 20)  (Qwen2.5: (16, 24, 24))
+mrope_interleaved     = True    (Qwen2.5: not interleaved)
+audio_token_id        = 151675  (Qwen2.5: 151646)
+image_token_id        = 151655
+video_token_id        = 151656
+audio_encoder         = qwen3_omni_moe_audio_encoder (Whisper-flavor, 128 mel)
+sampling_rate         = 16000
+```
+
+## To activate
+
+```bash
+source /opt/conda/etc/profile.d/conda.sh
+conda activate spur-qwen3
+```
+
+Existing Qwen2.5 / IV / AF3 work continues to use plain `spur` env unchanged.
+
+## Memory budget for 30B-A3B on 8× 40GB A100
+
+- BF16 weights ≈ 60 GB total. Single card cannot hold the full model.
+- ZeRO-3 with optimizer + param offload: ~60/8 ≈ 7.5 GB shard + ~10 GB activations + LoRA buffers ≈ 22 GB/card. Fits with margin.
+- LoRA targets `attention {q,k,v,o}_proj` only — expert MLPs (`SparseMoeBlock`) are NOT touched.
+- Router aux loss is disabled (`router_aux_loss_coef=0.0`, `output_router_logits=False`) so training signal is pure LM loss; matters because we only LoRA the attention path.
+
+## DeepSpeed ZeRO-3 integration status
+
+- `configs/ds_zero3_qwen3.json` is committed (CPU offload optimizer + param, BF16, auto-batch).
+- `train_spatial_beats_qa_qwen3.py` does **NOT** yet wire `deepspeed.initialize` — TODO. Until that's wired:
+  - **Stage1 projector_only on 8× 40GB**: try `BATCH_SIZE=1 GRAD_ACCUM_STEPS=8` with plain DDP. If the 30B forward OOMs in one card (likely), set `--device-map auto` via the trainer's `args.device_map` plumbing (single replica, no DDP) and drop NPROC=1.
+  - **Stage2/3 LoRA with optimizer state**: requires DS ZeRO-3 — add the `deepspeed.initialize(model, config_path=...)` call site in the trainer before that stage kicks off.
+- Expected work to wire DS into trainer (1–2 days): replace `optimizer.step()` / `loss.backward()` with `engine.step()` / `engine.backward(loss)`, replace `DistributedDataParallel` wrap with `engine`. The rest of the loop (data, valid, ckpt) is unchanged.
+

docs/qwen3_h20_ddp_curriculum.md

@@ -0,0 +1,281 @@
+# Qwen3-Omni H20 DDP 训练改动记录
+
+本文档记录将 Spatial-BEATs + Qwen3-Omni-30B-A3B-Instruct 训练从 8×A100-40G + `device_map=auto` 迁移到 8×H20-95G + 标准 DDP 的所有改动，包含：环境差异、代码修复、新增 shell 脚本，以及 easy → medium → hard 三难度课程化训练的入口。
+
+---
+
+## 1. 背景
+
+| 项目 | 旧环境 (spur-qwen3) | 新环境 (qwen3) |
+|------|--------------------|----------------|
+| GPU | 8 × A100 40 GB | 8 × H20 95 GB |
+| CUDA driver | 12.2 | 12.9 |
+| PyTorch | 2.10.0+cu128 | **2.9.0+cu128** |
+| Python | 3.10.19 | **3.12.13** |
+| transformers | 5.0.0.dev0（editable，远程 fork） | 5.0.0（site-packages） |
+| flash_attn | 未安装，使用 sdpa | **2.8.3，使用 flash_attention_2** |
+| accelerate | — | 1.13.0 |
+| deepspeed | 0.19.0 | 0.19.0 |
+| peft | 0.18.1 | 0.18.1 |
+
+关键差异：
+- **30B BF16 ≈ 60 GB**，旧 A100-40G 单卡不够 → 必须用 `device_map=auto`（accelerate 单进程多卡分片）。
+- 新 H20-95G **单卡可以容纳完整模型** → 可以走原生 DDP，每卡一份 replica，吞吐快 3-5×。
+- flash_attn 在新环境已可用 → 切到 `flash_attention_2` 进一步省显存、加速。
+
+---
+
+## 2. 关键路径
+
+| 用途 | 路径 |
+|------|------|
+| 项目根 | `/apdcephfs_fsgm/share_303840540/hunyuan/jensenwang/zzy/Spatial-Qwen` |
+| Qwen3 模型权重 | `/apdcephfs_fsgm/share_303840540/hunyuan/jensenwang/model_warehouse/Qwen3-Omni-30B-A3B-Instruct/`（66 GB，15 个 safetensors） |
+| transformers fork | `/apdcephfs_fsgm/share_303840540/hunyuan/jensenwang/zzy/transformers/src` |
+| BEATs ckpt | `/apdcephfs_cq10/share_1603164/.../spatial_beats_ov1_unified_v13d_exp/03_ov123_top4/best.pt` |
+| BEATs repo | `/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats` |
+| QA 数据根目录 | `/apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2_filtered_balanced/` |
+
+> ⚠️ QA 数据和 BEATs ckpt 在 `apdcephfs_cq10`（异地网络盘），跨区域 IO 比较慢。首次加载 train.jsonl 约 3 min，30B 模型 from_pretrained 约 15-20 min。
+
+### conda 环境激活
+
+```bash
+source /root/miniconda3/etc/profile.d/conda.sh
+conda activate qwen3
+```
+
+旧文档（`docs/qwen3_env.md`）里写的 `/opt/conda/etc/profile.d/conda.sh` 在本机不存在，conda 实际安装在 `/root/miniconda3`。
+
+---
+
+## 3. 代码改动
+
+### 3.1 `train_spatial_beats_qa_qwen3.py` —— 修复 DDP 模式下模型未上 GPU 的 bug
+
+**问题**：`_build_model_qwen3()` 只处理了 `device_map=auto` 的模块放置（去掉 accelerate hooks 后把 spatial 模块 pin 到 cuda:0）。当 `device_map=None`（DDP 模式）时，由于 `from_pretrained(low_cpu_mem_usage=True)` 把权重加载到 CPU，函数从未调用 `model.to(args.device)`，导致后续 `DDP(model, device_ids=[local_rank], ...)` 报 device 不匹配。Qwen2.5 的基线 `build_model()` 是有 `model.to(args.device)` 的，但 Qwen3 的 monkey-patch 替换没镜像这一分支。
+
+**修复**：在 `_build_model_qwen3()` 末尾、`return model` 之前添加：
+
+```python
+# DDP mode: move entire model to local GPU. ``from_pretrained`` loads to
+# CPU when ``low_cpu_mem_usage=True`` is set without a device_map; DDP
+# later wraps with device_ids=[local_rank] which requires the params to
+# already live on that device. (The base trainer's build_model does this
+# via ``model.to(args.device)``; the Qwen3 monkey-patched replacement
+# forgot to mirror that branch.)
+if device_map is None:
+    model.to(args.device)
+    _trainer.rank0_print(
+        f"[build_model_qwen3] DDP mode: moved model to {args.device}"
+    )
+```
+
+**验证**：smoke test 中日志出现 `[build_model_qwen3] DDP mode: moved model to cuda:0`，DDP wrap 成功，全部 8 ranks 进入训练循环，loss 在 1.0–3.0 之间正常下降。
+
+---
+
+### 3.2 `shell/launch_train_spatial_beats_qwen3.sh` —— 增加 `MAX_TRAIN_SAMPLES` / `MAX_VALID_SAMPLES` 转发
+
+底层 trainer (`train_spatial_beats_qa.py`) 本身支持 `--max-train-samples` / `--max-valid-samples` 这两个参数用于 smoke test，但旧的 shell 脚本没把环境变量透传过去。在 `common_args=(...)` 之后追加：
+
+```bash
+# Smoke / debug knobs: cap dataset size if requested (defaults: full split).
+if [[ -n "${MAX_TRAIN_SAMPLES:-}" ]]; then
+  common_args+=(--max-train-samples "${MAX_TRAIN_SAMPLES}")
+  echo "[config] max-train-samples = ${MAX_TRAIN_SAMPLES}"
+fi
+if [[ -n "${MAX_VALID_SAMPLES:-}" ]]; then
+  common_args+=(--max-valid-samples "${MAX_VALID_SAMPLES}")
+  echo "[config] max-valid-samples = ${MAX_VALID_SAMPLES}"
+fi
+```
+
+这样：
+```bash
+MAX_TRAIN_SAMPLES=64 MAX_VALID_SAMPLES=8 bash shell/launch_train_spatial_beats_qwen3.sh
+```
+就能跑一个真实的小批数据 smoke。
+
+---
+
+## 4. 新增脚本
+
+### 4.1 `shell/launch_train_spatial_beats_qwen3_h20_ddp.sh` —— H20 DDP 入口
+
+H20 95 GB 单卡能容纳完整 30B 模型，因此切换到原生 DDP（一进程一卡 replica）。脚本是对 `launch_train_spatial_beats_qwen3.sh` 的薄封装，改动如下：
+
+```bash
+unset DEVICE_MAP                              # 触发底层脚本的 torchrun DDP 分支
+export ATTN_IMPL="${ATTN_IMPL:-flash_attention_2}"   # qwen3 env 已装 flash_attn 2.8.3
+# 每卡 BS（保守起步，stage1 实测 ~67 GB / 95 GB）：
+export STAGE1_BATCH_SIZE="${STAGE1_BATCH_SIZE:-4}"   # projector_only：最省显存
+export STAGE2_BATCH_SIZE="${STAGE2_BATCH_SIZE:-2}"   # encoder_lora：加 LoRA optimizer
+export STAGE3_BATCH_SIZE="${STAGE3_BATCH_SIZE:-1}"   # beats_lora：解冻 BEATs 激活更大
+# 全局 batch 通过 grad_accum 维持在 64：
+export STAGE1_GRAD_ACCUM="${STAGE1_GRAD_ACCUM:-2}"   # 4*2*8 = 64
+export STAGE2_GRAD_ACCUM="${STAGE2_GRAD_ACCUM:-4}"   # 2*4*8 = 64
+export STAGE3_GRAD_ACCUM="${STAGE3_GRAD_ACCUM:-8}"   # 1*8*8 = 64
+export SPATIAL_QWEN_NCCL_TIMEOUT_MIN="${SPATIAL_QWEN_NCCL_TIMEOUT_MIN:-120}"
+exec bash "${ROOT_DIR}/shell/launch_train_spatial_beats_qwen3.sh" "$@"
+```
+
+注意：所有变量都用 `${VAR:-default}` 风格，可被外层覆盖。
+
+> 实测显存（stage1 projector_only, BS=1, GC 开启）：约 **66-69 GB / 卡**，剩 ~25 GB 给激活/通信。BS=4 大概率能放下；如果还有富余可以试 BS=8。
+
+---
+
+### 4.2 `shell/launch_train_curriculum_qwen3_h20_ddp.sh` —— easy/medium/hard 课程化训练
+
+按用户要求实现 **easy 三阶段 → medium stage3 续训 → hard stage3 续训** 的课程：
+
+| 阶段 | QA 数据 | 训练模式 | 起点 |
+|------|---------|----------|------|
+| Phase 1 — easy | `easy/` (627 K train, 66 K valid) | stage1→stage2→stage3 | 从头 |
+| Phase 2 — medium | `medium_plus_easy10/` (607 K train, 65 K valid) | 仅 stage3 (beats_lora) | resume from easy stage3 best |
+| Phase 3 — hard | `hard_plus_med20_easy10/` (377 K train, 40 K valid) | 仅 stage3 (beats_lora) | resume from medium stage3 best |
+
+> `medium_plus_easy10` 表示 medium 主体 + 10% easy 回放，`hard_plus_med20_easy10` 表示 hard + 20% medium + 10% easy 回放。**纯 hard 目录只有 test.jsonl，没有 train/valid，所以 hard 阶段必须用 `_plus_*` 课程版本。**
+
+#### 关键设计
+
+- **不重写底层逻辑**：每个 phase 都是一次 `bash launch_train_spatial_beats_qwen3_h20_ddp.sh` 的调用，通过 env 变量传 `QA_ROOT` / `RUN_ROOT` / `START_STAGE` / `STAGE3_RESUME_CKPT` / `STAGE3_RESUME_MODEL_ONLY=1`，让现有 stage 切换逻辑负责 LoRA 装配 + ckpt 加载。
+- **medium / hard 用更小学习率**：避免覆盖 easy 阶段已学到的 LoRA + projector + BEATs 权重。默认：
+  ```
+  MEDIUM_STAGE3_LR=2e-5  MEDIUM_STAGE3_LORA_LR=2e-5
+  MEDIUM_STAGE3_PROJECTOR_LR=1e-6  MEDIUM_STAGE3_BEATS_LR=5e-7
+  HARD_STAGE3_LR=1e-5  HARD_STAGE3_LORA_LR=1e-5
+  HARD_STAGE3_PROJECTOR_LR=1e-6  HARD_STAGE3_BEATS_LR=5e-7
+  ```
+- **依赖检查**：每个 phase 进入前检查对应 jsonl 存在；medium/hard 还会检查上一 phase 的 best_trainable.pt 是否就绪，缺失则 fail-fast。
+- **可跳过任意 phase**：`SKIP_EASY=1` / `SKIP_MEDIUM=1` / `SKIP_HARD=1`；easy 内部可用 `EASY_START_STAGE=2` 等续训。
+
+#### 输出目录结构
+
+```
+runs/qwen3_curriculum/
+  easy/
+    stage1_projector/checkpoints/{best,last,epoch_NNN}_trainable.pt
+    stage2_encoder_lora/checkpoints/...
+    stage3_beats_lora/checkpoints/best_trainable.pt   # → medium 起点
+  medium/
+    stage3_beats_lora/checkpoints/best_trainable.pt   # → hard 起点
+  hard/
+    stage3_beats_lora/checkpoints/best_trainable.pt
+```
+
+---
+
+## 5. 启动命令
+
+### 5.1 释放 GPU（之前残留的 vllm / torchrun 进程）
+
+```bash
+# 终止占用 8 卡的 ms-swift vllm deploy
+kill -9 1194399 1194463 1194590 1194717 1194781 1194971 1195098 1195165
+pkill -9 -f 'torchrun.*train_spatial' || true
+pkill -9 -f 'train_spatial_beats_qa_qwen3' || true
+nvidia-smi   # 确认 8 卡显存清空
+```
+
+### 5.2 启动完整课程训练
+
+建议放进 tmux/screen，预计运行多天。
+
+```bash
+source /root/miniconda3/etc/profile.d/conda.sh
+conda activate qwen3
+cd /apdcephfs_fsgm/share_303840540/hunyuan/jensenwang/zzy/Spatial-Qwen
+
+mkdir -p logs
+GPUS=0,1,2,3,4,5,6,7 \
+MODEL_ID=/apdcephfs_fsgm/share_303840540/hunyuan/jensenwang/model_warehouse/Qwen3-Omni-30B-A3B-Instruct \
+QWEN3_OMNI_FORK=/apdcephfs_fsgm/share_303840540/hunyuan/jensenwang/zzy/transformers/src \
+RUN_ROOT_BASE=./runs/qwen3_curriculum \
+bash shell/launch_train_curriculum_qwen3_h20_ddp.sh 2>&1 \
+  | tee logs/qwen3_curriculum_$(date +%Y%m%d_%H%M).log
+```
+
+### 5.3 常用 override
+
+```bash
+# 跳过部分 phase（已经跑完 easy 想继续 medium）
+SKIP_EASY=1 bash shell/launch_train_curriculum_qwen3_h20_ddp.sh
+
+# 只跑 hard
+SKIP_EASY=1 SKIP_MEDIUM=1 bash shell/launch_train_curriculum_qwen3_h20_ddp.sh
+
+# easy 从 stage2 续（stage1 已完成）
+EASY_START_STAGE=2 bash shell/launch_train_curriculum_qwen3_h20_ddp.sh
+
+# 调整 BS（实测 stage1 还有大量显存余量时尝试）
+STAGE1_BATCH_SIZE=8 STAGE1_GRAD_ACCUM=1 \
+  bash shell/launch_train_curriculum_qwen3_h20_ddp.sh
+
+# 调整 epoch 数
+STAGE1_EPOCHS=2 STAGE2_EPOCHS=3 STAGE3_EPOCHS=3 \
+MEDIUM_STAGE3_EPOCHS=2 HARD_STAGE3_EPOCHS=2 \
+  bash shell/launch_train_curriculum_qwen3_h20_ddp.sh
+
+# Smoke：每 phase 32 训练样本 + 1 epoch（验证 wiring，不训练）
+MAX_TRAIN_SAMPLES=32 MAX_VALID_SAMPLES=8 \
+STAGE1_EPOCHS=1 STAGE2_EPOCHS=1 STAGE3_EPOCHS=1 \
+MEDIUM_STAGE3_EPOCHS=1 HARD_STAGE3_EPOCHS=1 \
+RUN_ROOT_BASE=./runs/qwen3_curriculum_smoke \
+  bash shell/launch_train_curriculum_qwen3_h20_ddp.sh
+```
+
+---
+
+## 6. 已验证 / 已知行为
+
+### 6.1 已验证
+
+- **环境**：qwen3 conda env 内 torch 2.9.0+cu128、transformers 5.0.0、flash_attn 2.8.3、peft、accelerate、tensorboard、soundfile 全部 import 通过。
+- **fork 可达**：`/apdcephfs_fsgm/.../zzy/transformers/src` 下 `qwen3_omni_moe` 子模块可正常 import；config.json 解析后 `hidden_size=2048`、`audio_token_id=151675` 与 `docs/qwen3_env.md` 描述一致。
+- **DDP 启动**：8 ranks 通过 torchrun 正常初始化，`from_pretrained` 后通过新加的 `model.to(args.device)` 上 GPU，DDP wrap（`find_unused_parameters=True`）成功。
+- **训练循环**：stage1 BS=1 项目下 ~2 it/s，loss 1.0–3.0 区间合理震荡。
+- **显存（stage1 projector_only, BS=1, GC=on）**：~66-69 GB / 卡，95 GB 余 ~25 GB，BS=4 安全可行。
+
+### 6.2 已知 quirk
+
+- **跨区域 NFS 慢**：每次 from_pretrained 加载 30B 权重约 15-20 min，是当前最大的启动开销。如果反复实验值得做：
+  ```bash
+  # 把模型 cp 到本地 ssd（如果有）
+  rsync -av /apdcephfs_fsgm/.../Qwen3-Omni-30B-A3B-Instruct/ /local/ssd/Qwen3-Omni-30B-A3B-Instruct/
+  ```
+- **DDP find_unused_parameters=True 警告**：基底 trainer 出于稳健性默认开启，每个 forward 多遍历一次 autograd 图。stage1 实测对吞吐影响不大，stage2/3 LoRA 路径才真正需要它，保持现状。
+- **`use_cache=True` 与 GC 冲突**：transformers 会自动回退到 `use_cache=False`，仅有警告，可忽略。
+- **vllm 占卡**：本机此前长时间运行 8 个 ms-swift vllm deploy（端口 8100-8107，`Qwen3omni-thinking-opsd-merged-ablation1500`），训练前必须 `kill -9` 释放。
+
+### 6.3 ⚠️ 未验证但需要关注
+
+- **BS=4 / 8 / 16 的真实显存峰值**：smoke 没跑到稳态，启动时 GPU 还有 vllm 残留导致一次 OOM。建议第一次跑真实训练前用 `nvidia-smi --query-gpu=memory.used --format=csv -l 5` 监控前 100 步峰值，按需下调 BS。
+- **stage3 解冻 BEATs 的显存**：BEATs encoder backbone 进入训练梯度，且会调 `torch.stft` / `cuFFT`，旧文档强调要小心 `cufft_plan_cache` 碎片（基底 trainer 已经把 `max_size=8` + `empty_cache(每500 step)`）。stage3 BS=1 应该安全，但请观察前几个 epoch 是否稳定。
+- **medium/hard resume 的 `find_unused_parameters`**：medium/hard 从 easy stage3 best 续训，如果 LoRA target 解析顺序在不同 stage 略有不同，可能有未使用参数；已默认开启 `find_unused_parameters=True`，无需手动处理。
+
+---
+
+## 7. 文件改动清单
+
+| 路径 | 改动 |
+|------|------|
+| `train_spatial_beats_qa_qwen3.py` | **修改** —— `_build_model_qwen3()` 末尾追加 `if device_map is None: model.to(args.device)`（约 7 行新增） |
+| `shell/launch_train_spatial_beats_qwen3.sh` | **修改** —— 在 `common_args` 后追加 MAX_TRAIN_SAMPLES / MAX_VALID_SAMPLES 转发（约 8 行新增） |
+| `shell/launch_train_spatial_beats_qwen3_h20_ddp.sh` | **新增** —— H20 DDP 入口（~70 行） |
+| `shell/launch_train_curriculum_qwen3_h20_ddp.sh` | **新增** —— easy/medium/hard 课程化主控脚本（~140 行） |
+| `docs/qwen3_h20_ddp_curriculum.md` | **新增** —— 本文档 |
+
+未改动（按原约束保留）：
+- `train_spatial_beats_qa.py`（Qwen2.5 基底 trainer）
+- `transformers` fork 内任何文件
+- `spatial_qwen/model/` 与 `spatial_qwen/modules/` 任何文件
+
+---
+
+## 8. 后续 TODO（不阻塞当前训练）
+
+- **DeepSpeed ZeRO-3 接入**：`configs/ds_zero3_qwen3.json` 已存在，但 trainer 还没接 `deepspeed.initialize`。当前 H20 DDP 已经够用，仅在多机扩展或后续更大模型时再做。
+- **本地 SSD 缓存模型权重**：消除每次 15-20 min 的跨区域 NFS 启动开销。
+- **课程化数据加权策略实验**：当前 medium / hard 直接复用预生成的 `_plus_*` jsonl。后续可以尝试在 trainer 里做 on-the-fly 难度采样。

docs/seld233_implementation.md

@@ -0,0 +1,417 @@
+# SELD233 Spatial Implementation Guide
+
+## Purpose
+This document is the implementation guide for the independent SELD233 spatial modality.
+
+It answers four questions:
+- which classes and functions were implemented for the core path
+- which classes are already scaffolded and should be reused as-is
+- what each interface accepts and returns
+- what concrete code is still deferred outside the current scope
+
+This guide should be read together with:
+- `docs/spatial_encoder_plan.md`
+- `docs/seld233_spatial_scaffold.md`
+
+## Fixed Constraints
+- Do not modify the DCASE baseline repository.
+- Do not modify the original QwenOmni main classes for this feature; use the new spatial subclasses.
+- A batch must be either all mono or all FOA.
+- Mono batch:
+  - must contain exactly one `<|AUDIO|>`
+  - must not contain `<|spatial|>`
+  - only the original audio path is active
+- FOA batch:
+  - must contain exactly one `<|AUDIO|>`
+  - must contain exactly one `<|spatial|>`
+  - both audio and spatial paths are active
+- Audio is clipped to `0-20s`.
+- Sample rate is `16kHz`.
+- Spatial token rate target is `2.5 Hz`.
+- Full `20s` limits:
+  - max samples: `320000`
+  - max feature frames: `1000`
+  - max SELD frames: `200`
+  - max spatial tokens: `50`
+
+## Status Summary
+
+### Already scaffolded and reusable
+- `qwen2_5_omni_spur/spatial_seld233_utils.py`
+- `qwen2_5_omni_spur/modules/seldnet233_spatial_adapter.py`
+- `qwen2_5_omni_spur/modules/spatial_token_projector.py`
+- `qwen2_5_omni_spur/processing_qwen2_5_omni_spatial.py`
+- `qwen2_5_omni_spur/modeling_qwen2_5_omni_spatial.py`
+- config additions in `qwen2_5_omni_spur/configuration_qwen2_5_omni.py`
+
+### Must still be implemented
+- training script migration to the new spatial subclasses
+- inference script migration to the new spatial subclasses
+
+### Explicitly deferred for now
+- `image + spatial` joint RoPE
+- `use_audio_in_video=True`
+- `return_audio=True` talker path with spatial modality
+
+## Implementation Inventory
+
+| Priority | File | Symbol | Current State | Required Action |
+| --- | --- | --- | --- | --- |
+| P0 | `qwen2_5_omni_spur/modules/seldnet233_feature_bridge.py` | `SeldNet233FeatureBridge._extract_online_features` | implemented | no further action for the current scope |
+| P0 | `qwen2_5_omni_spur/modules/seldnet233_backbone.py` | `SeldNet233Backbone._run_seldnet_backbone` | implemented | no further action for the current scope |
+| P1 | `training-qwen-omni/train_spur_spatial_only_hf.py` | `main()` and config wiring | legacy | migrate only if this script is needed for training runs |
+| P1 | `qwen-omni-inference.py` | script body | legacy | migrate only if this script is needed for ad hoc inference |
+| P2 | `qwen2_5_omni_spur/modeling_qwen2_5_omni_spatial.py` | `get_rope_index` | partial | only needed later if image + spatial is required |
+| P2 | `qwen2_5_omni_spur/modeling_qwen2_5_omni_spatial.py` | `generate(return_audio=True)` path | partial | only needed if talker output is required later |
+
+## Detailed Interface Guide
+
+### 1. `SeldNet233FeatureBridge`
+File:
+- `qwen2_5_omni_spur/modules/seldnet233_feature_bridge.py`
+
+Class:
+- `SeldNet233FeatureBridge`
+
+Already implemented:
+- constructor
+- input validation
+- waveform mask -> valid length conversion
+- waveform length -> feature length conversion
+- output mask bookkeeping
+- online baseline-aligned STFT / log-mel / intensity-vector extraction
+- baseline `foa_wts` normalization
+
+Input contract:
+- `spatial_audio`: `torch.Tensor`, shape `[B, T_audio_max, 4]`
+- `spatial_audio_lengths`: `torch.LongTensor`, shape `[B]`
+- `feature_lengths`: `torch.LongTensor`, shape `[B]`
+- `feature_attention_mask`: `torch.BoolTensor`, shape `[B, T_feat_max]`
+
+Output contract:
+- `SeldNet233FeatureBridgeOutput.features`: `[B, 7, T_feat_max, 64]`
+- `SeldNet233FeatureBridgeOutput.feature_attention_mask`: `[B, T_feat_max]`
+- `SeldNet233FeatureBridgeOutput.feature_lengths`: `[B]`
+
+Required implementation content:
+1. Load or derive the exact task-233 feature configuration.
+2. Reproduce the baseline feature math for FOA:
+   - `4ch` log-mel
+   - `3ch` FOA intensity vector
+3. Use the task-233 normalization weights.
+4. Keep output frame count aligned with `feature_lengths`.
+5. Zero-fill padded tail frames only after valid frames are computed.
+
+Recommended private helpers to add inside this file:
+- `_load_task233_feature_config()`
+- `_load_feature_stats()`
+- `_compute_stft(...)`
+- `_compute_logmel_channels(...)`
+- `_compute_intensity_vector_channels(...)`
+- `_normalize_feature_tensor(...)`
+
+Important shape transitions:
+- raw FOA waveform: `[B, T_audio_max, 4]`
+- internal channels-first view: `[B, 4, T_audio_max]`
+- mel-like features before stacking: `[B, 4, T_feat_max, 64]`
+- intensity-vector features: `[B, 3, T_feat_max, 64]`
+- final baseline tensor: `[B, 7, T_feat_max, 64]`
+
+Validation checklist:
+- same sample length gives same `T_feat` as `samples_to_feature_frames`
+- values match offline baseline extraction on the same clip
+- normalization path is identical to baseline task `233`
+
+### 2. `SeldNet233Backbone`
+File:
+- `qwen2_5_omni_spur/modules/seldnet233_backbone.py`
+
+Class:
+- `SeldNet233Backbone`
+
+Already implemented:
+- constructor
+- input validation
+- feature length -> SELD length conversion
+- output mask bookkeeping
+- dynamic baseline loading
+- checkpoint restore with shape-compatible filtering
+- last MHSA LayerNorm hidden capture
+
+Input contract:
+- `seld233_features`: `torch.Tensor`, shape `[B, 7, T_feat_max, 64]`
+- `seld233_feature_lengths`: `torch.LongTensor`, shape `[B]`
+- `hidden_lengths`: `torch.LongTensor`, shape `[B]`
+- `hidden_attention_mask`: `torch.BoolTensor`, shape `[B, T_seld_max]`
+
+Output contract:
+- `SeldNet233BackboneOutput.hidden_states`: `[B, T_seld_max, 128]`
+- `SeldNet233BackboneOutput.hidden_attention_mask`: `[B, T_seld_max]`
+- `SeldNet233BackboneOutput.hidden_lengths`: `[B]`
+
+Required implementation content:
+1. Dynamically load `parameters.py` and `seldnet_model.py` from the baseline repo path.
+2. Build the task-233 baseline model.
+3. Restore the pretrained checkpoint from `seld233_checkpoint_path`.
+4. Freeze the backbone if `seld233_freeze_backbone=True`.
+5. Register a hook at the final MHSA shared representation.
+6. Feed the baseline-compatible features into the model.
+7. Return the captured hidden tensor, not accdoa heads and not sed logits.
+
+Recommended private helpers to add inside this file:
+- `_load_baseline_modules()`
+- `_build_seld233_model()`
+- `_load_checkpoint_weights()`
+- `_register_hidden_hook()`
+- `_prepare_baseline_input_layout(...)`
+
+Important shape transitions:
+- model input for this wrapper: `[B, 7, T_feat_max, 64]`
+- likely baseline internal flattened layout: `[B, T_feat_max, 448]`
+- captured MHSA hidden: `[B, T_seld_max, 128]`
+
+Validation checklist:
+- checkpoint path exists
+- hook fires exactly once per forward
+- hidden dim is exactly `128`
+- hidden length matches `feature_frames_to_seld_frames`
+
+### 3. `SeldNet233SpatialAdapter`
+File:
+- `qwen2_5_omni_spur/modules/seldnet233_spatial_adapter.py`
+
+Class:
+- `SeldNet233SpatialAdapter`
+
+Status:
+- already usable once feature bridge and backbone are implemented
+
+Input modes already supported:
+- raw audio path:
+  - `spatial_audio [B, T_audio_max, 4]`
+- offline feature path:
+  - `seld233_features [B, 7, T_feat_max, 64]`
+- direct hidden path:
+  - `seld233_hidden_states [B, T_seld_max, 128]`
+
+Output:
+- `spatial_tokens [B, T_spat_max, 256]`
+- `spatial_token_attention_mask [B, T_spat_max]`
+- `spatial_token_lengths [B]`
+
+No new implementation is required here unless you want to change:
+- downsampling rule
+- token MLP architecture
+- token dimension
+
+Current downsampling rule:
+- `T_spat = ceil(T_seld / 4)`
+- effective rate: `10 Hz -> 2.5 Hz`
+
+### 4. `SpatialTokenProjector`
+File:
+- `qwen2_5_omni_spur/modules/spatial_token_projector.py`
+
+Class:
+- `SpatialTokenProjector`
+
+Status:
+- already implemented
+
+Input:
+- `spatial_tokens [B, T_spat, D_in]`
+
+Output:
+- projected tokens `[B, T_spat, D_llm]`
+
+No further code is required here unless model capacity needs tuning.
+
+### 5. `Qwen2_5OmniSpatialProcessor`
+File:
+- `qwen2_5_omni_spur/processing_qwen2_5_omni_spatial.py`
+
+Class:
+- `Qwen2_5OmniSpatialProcessor`
+
+Status:
+- implemented scaffold
+- should be reused, not rewritten
+
+Important public interfaces:
+- `__call__(...)`
+- `sync_spatial_tokenizer_with_model(model)`
+
+Input behavior:
+- mono batch:
+  - prompt must include one `<|AUDIO|>`
+  - prompt must not include `<|spatial|>`
+- FOA batch:
+  - prompt must include one `<|AUDIO|>`
+  - prompt must include one `<|spatial|>`
+
+Output keys already supported:
+- `input_ids`
+- `input_features`
+- `feature_attention_mask`
+- `spatial_audio`
+- `spatial_audio_attention_mask`
+- `spatial_audio_lengths`
+- `spatial_tokens`
+- `seld233_features`
+- `seld233_feature_attention_mask`
+- `seld233_feature_lengths`
+- `spatial_token_lengths`
+
+No core algorithm is missing here.
+
+What still needs to happen outside this file:
+- training script must actually instantiate this processor
+- inference script must actually instantiate this processor
+- both scripts must call `processor.sync_spatial_tokenizer_with_model(model)`
+
+### 6. `Qwen2_5OmniSpatialThinkerForConditionalGeneration`
+File:
+- `qwen2_5_omni_spur/modeling_qwen2_5_omni_spatial.py`
+
+Class:
+- `Qwen2_5OmniSpatialThinkerForConditionalGeneration`
+
+Status:
+- spatial injection path is implemented
+- tokenizer sync helper is implemented
+- prefill spatial routing is implemented
+- video + audio + spatial RoPE is implemented for `use_audio_in_video=False`
+
+Important public interfaces:
+- `forward(...)`
+- `prepare_inputs_for_generation(...)`
+- `sync_spatial_tokenizer(...)`
+
+Input contract:
+- normal Qwen text/audio inputs
+- optional spatial inputs:
+  - `spatial_audio [B, T_audio_max, 4]`
+  - `seld233_features [B, 7, T_feat_max, 64]`
+  - `spatial_tokens [B, T_spat_max, D_spat]`
+  - matching length or mask tensors
+
+Output behavior:
+- if no spatial input is provided, it falls back to the base thinker
+- if spatial input is provided, it injects projected spatial embeddings at
+  `<|spatial|>` positions
+
+Still missing only if scope expands later:
+- true multimodal RoPE for `image + spatial`
+- talker/audio-output path
+
+### 7. `Qwen2_5OmniSpatialForConditionalGeneration`
+File:
+- `qwen2_5_omni_spur/modeling_qwen2_5_omni_spatial.py`
+
+Class:
+- `Qwen2_5OmniSpatialForConditionalGeneration`
+
+Status:
+- implemented scaffold
+
+Important public interfaces:
+- `sync_spatial_tokenizer(...)`
+- `generate(...)`
+
+No core code is missing for text-only generation with audio/spatial input.
+
+Still deferred:
+- `return_audio=True` path
+
+## Script-Level Work
+
+### 8. Training Script
+File:
+- `training-qwen-omni/train_spur_spatial_only_hf.py`
+
+Current issue:
+- it still uses the old SPUR audio-fusion line
+
+Minimum required changes:
+1. Replace imports:
+   - old:
+     - `Qwen2_5OmniProcessor`
+     - `Qwen2_5OmniForConditionalGeneration`
+   - new:
+     - `Qwen2_5OmniSpatialProcessor`
+     - `Qwen2_5OmniSpatialForConditionalGeneration`
+2. Replace old SPUR config overrides with new `seld233_*` thinker config fields.
+3. After loading processor and model, call:
+   - `processor.sync_spatial_tokenizer_with_model(model)`
+4. Update prompt construction:
+   - mono batch prompt: includes `<|AUDIO|>` only
+   - FOA batch prompt: includes `<|AUDIO|><|spatial|>`
+5. Ensure collator preserves mono or FOA batch homogeneity.
+6. Ensure audio clips are clipped to `20s`.
+7. Ensure training batch uses the new processor output keys.
+
+Functions that most likely need modification:
+- `apply_spur_config_overrides(...)`
+- `main()`
+
+### 9. Inference Script
+File:
+- `qwen-omni-inference.py`
+
+Current issue:
+- it still uses the old processor/model path
+
+Minimum required changes:
+1. Replace imports with the new spatial subclasses.
+2. Load FOA audio without collapsing channels.
+3. Build prompt with canonical tokens:
+   - FOA: `<|AUDIO|><|spatial|> ...`
+   - mono: `<|AUDIO|> ...`
+4. Call:
+   - `processor.sync_spatial_tokenizer_with_model(model)`
+5. Print or assert:
+   - `spatial_token_lengths`
+   - number of `<|spatial|>` token positions after tokenization
+6. Run one prefill/generation sanity check.
+
+There is no reusable function boundary in the current script; the top-level
+script body itself needs to be migrated.
+
+## Suggested Implementation Order
+1. Implement `SeldNet233FeatureBridge._extract_online_features`.
+2. Implement `SeldNet233Backbone._run_seldnet_backbone`.
+3. Run a direct module test:
+   - `spatial_audio -> feature bridge -> backbone -> adapter`
+4. Migrate `training-qwen-omni/train_spur_spatial_only_hf.py`.
+5. Migrate `qwen-omni-inference.py`.
+6. Only after that, consider optional RoPE/talker extensions.
+
+## Required Checks After Each Stage
+
+### After feature bridge
+- same clip gives matching online vs offline task-233 features
+- feature tensor is `[B, 7, T_feat_max, 64]`
+- feature mask matches `feature_lengths`
+
+### After backbone
+- checkpoint loads successfully
+- captured hidden is `[B, T_seld_max, 128]`
+- hidden length matches derived `T_seld`
+
+### After end-to-end spatial adapter
+- `spatial_tokens` is `[B, T_spat_max, 256]`
+- `spatial_token_lengths` equals `ceil(T_seld / 4)`
+- no placeholder count mismatch
+
+### After training/inference migration
+- `<|spatial|>` is present in tokenizer vocab
+- `config.thinker_config.spatial_token_index` is set
+- thinker embedding size matches tokenizer size
+- mono batch runs without spatial injection
+- FOA batch runs with spatial injection
+
+## Out of Scope For The Next Coding Stage
+- support mixed mono/FOA batches
+- support image/video + spatial at the same time
+- support talker audio generation with spatial tokens
+- change the task-233 feature definition or baseline architecture

docs/seld233_scaffold.md

@@ -0,0 +1,224 @@
+# SELD233 Spatial Scaffold
+
+## Purpose
+This scaffold adds a new, non-invasive code path for the independent SELD233 spatial modality without changing the existing QwenOmni audio / image / video implementation.
+
+The scaffold focuses on:
+- public class boundaries
+- tensor contracts
+- length bookkeeping
+- placeholder expansion
+- thinker-side spatial injection wiring
+
+The current state is no longer scaffold-only for the core `audio + spatial + text`
+path:
+- online `FOA -> 7ch` feature extraction is implemented
+- SELD backbone loading + checkpoint restore + MHSA hidden capture is implemented
+- the remaining deferred items are limited to `image + spatial`,
+  `use_audio_in_video=True`, and `return_audio=True`
+
+## New Files
+
+### `qwen2_5_omni_spur/spatial_seld233_utils.py`
+Shared length and mask helpers.
+
+Key utilities:
+- `attention_mask_to_lengths(attention_mask) -> [B]`
+- `build_1d_attention_mask(lengths, max_length) -> [B, T]`
+- `samples_to_feature_frames(num_samples) -> [B]`
+- `feature_frames_to_seld_frames(num_feature_frames) -> [B]`
+- `seld_frames_to_spatial_tokens(num_seld_frames) -> [B]`
+- `samples_to_spatial_length_bundle(num_samples) -> Seld233LengthBundle`
+
+### `qwen2_5_omni_spur/modules/seldnet233_feature_bridge.py`
+Online feature bridge scaffold.
+
+Input:
+- `spatial_audio`: `[B, T_audio, 4]`
+- `spatial_audio_attention_mask`: `[B, T_audio]`
+- `spatial_audio_lengths`: `[B]`
+
+Expected output:
+- `features`: `[B, 7, T_feat_max, 64]`
+- `feature_attention_mask`: `[B, T_feat_max]`
+- `feature_lengths`: `[B]`
+
+Current status:
+- input validation implemented
+- length bookkeeping implemented
+- baseline-aligned STFT / mel / intensity-vector extraction implemented
+- baseline `foa_wts` normalization implemented
+
+### `qwen2_5_omni_spur/modules/seldnet233_backbone.py`
+SELD backbone scaffold.
+
+Input:
+- `seld233_features`: `[B, 7, T_feat_max, 64]`
+- `seld233_feature_attention_mask`: `[B, T_feat_max]`
+- `seld233_feature_lengths`: `[B]`
+
+Expected output:
+- `hidden_states`: `[B, T_seld_max, 128]`
+- `hidden_attention_mask`: `[B, T_seld_max]`
+- `hidden_lengths`: `[B]`
+
+Current status:
+- input validation implemented
+- feature-to-SELD length mapping implemented
+- dynamic baseline loading implemented
+- checkpoint restore implemented
+- last-MHSA hidden capture implemented
+
+### `qwen2_5_omni_spur/modules/spatial_token_projector.py`
+Implemented MLP projector.
+
+Input:
+- `spatial_tokens`: `[B, T_spat, D_in]`
+
+Output:
+- projected tokens: `[B, T_spat, D_out]`
+
+### `qwen2_5_omni_spur/modules/seldnet233_spatial_adapter.py`
+Implemented low-rate token adapter around the bridge/backbone.
+
+Supported input modes:
+1. `spatial_audio`
+2. `seld233_features`
+3. `seld233_hidden_states`
+
+Main output:
+- `spatial_tokens`: `[B, T_spat_max, 256]`
+- `spatial_token_attention_mask`: `[B, T_spat_max]`
+- `spatial_token_lengths`: `[B]`
+
+Implemented:
+- mode routing
+- hidden-state downsampling from `10 Hz` to `2.5 Hz`
+- token MLP head
+- upstream `spatial_audio -> features -> hidden_states` path is now implemented
+
+### `qwen2_5_omni_spur/processing_qwen2_5_omni_spatial.py`
+New processor subclass: `Qwen2_5OmniSpatialProcessor`
+
+What it does:
+1. accepts mono or FOA audio and normalizes it to `[C, T_audio]`
+2. truncates every sample to `0-20s`, with explicit warnings on clipping
+3. rejects mixed `mono + FOA` batches up front
+4. keeps the original audio path aligned to the same clipped window
+5. normalizes prompt aliases by converting `<|audio|>` to the canonical
+   tokenizer token `<|AUDIO|>`
+6. for FOA samples only, creates:
+   - `spatial_audio`: `[B, 320000, 4]`
+   - `spatial_audio_attention_mask`: `[B, 320000]`
+   - `spatial_audio_lengths`: `[B]`
+   - `spatial_token_lengths`: `[B]`
+7. supports explicit processor-side spatial inputs:
+   - `spatial_tokens`
+   - `seld233_features`
+   - `seld233_feature_attention_mask`
+   - `seld233_feature_lengths`
+   - `spatial_token_lengths` only as a companion to `spatial_tokens` or
+     `seld233_features`, never as a standalone field
+8. validates prompt constraints:
+   - every sample must contain exactly one `<|AUDIO|>`
+   - FOA samples must contain exactly one `<|spatial|>`
+   - mono samples must contain zero `<|spatial|>`
+9. expands one `<|spatial|>` placeholder into repeated `<|spatial|>` tokens
+10. delegates normal audio / image / video handling to the base processor
+
+Current status:
+- fully wired for placeholder expansion, mono/FOA branching, tensor packaging,
+  and processor-side spatial fallback inputs
+- registers `<|spatial|>` into the tokenizer
+- exposes `sync_spatial_tokenizer_with_model(model)` to synchronize tokenizer
+  vocab growth with model embedding resize
+
+### `qwen2_5_omni_spur/modeling_qwen2_5_omni_spatial.py`
+New model subclasses:
+- `Qwen2_5OmniSpatialThinkerForConditionalGeneration`
+- `Qwen2_5OmniSpatialForConditionalGeneration`
+
+Thinker behavior:
+1. reuse original audio/image/video merge behavior
+2. resolve `spatial_tokens` from:
+   - direct input, or
+   - spatial adapter
+3. project tokens to LLM hidden size
+4. pack only valid spatial rows across the batch
+5. replace `<|spatial|>` positions using `masked_scatter`
+
+Current status:
+- precomputed `spatial_tokens` path is wired
+- raw-audio / raw-feature path is wired up to the adapter
+- variable-length `B>1` spatial-token injection is now packed correctly before
+  `masked_scatter`
+- online feature bridge + SELD backbone internals are implemented
+- RoPE now supports:
+  - `text + audio + spatial`
+  - `text + video + audio + spatial`
+  with fixed modal order `<|VIDEO|><|AUDIO|><|spatial|>` and
+  `use_audio_in_video=False`
+- image + spatial remains unsupported
+- `return_audio=True` with spatial currently raises `NotImplementedError`
+
+## Minimal Config Fields Added
+Added to `Qwen2_5OmniThinkerConfig`:
+- `spatial_token_index`
+- `use_seld233_spatial_modality`
+- `seld233_checkpoint_path`
+- `seld233_baseline_repo_path`
+- `seld233_task_id`
+- `seld233_feature_stats_dir`
+- `seld233_num_feature_channels`
+- `seld233_num_mel_bins`
+- `seld233_encoder_dim`
+- `seld233_token_dim`
+- `seld233_token_rate_hz`
+- `seld233_downsample_factor`
+- `seld233_projector_hidden_dim`
+- `seld233_freeze_backbone`
+- `seld233_max_audio_seconds`
+
+## Current Call Chain
+For the intended text+FOA(+spatial) path:
+
+1. `Qwen2_5OmniSpatialProcessor.__call__`
+2. base processor handles original audio path
+3. processor adds spatial tensors and expands `<|spatial|>`
+4. `Qwen2_5OmniSpatialForConditionalGeneration.generate(...)`
+5. `Qwen2_5OmniSpatialThinkerForConditionalGeneration.forward(...)`
+6. thinker resolves spatial tokens
+7. thinker projects spatial tokens, packs valid rows, and injects them into `inputs_embeds`
+8. original thinker forward continues for text/audio merging
+
+For the intended text+video+FOA+spatial path:
+
+1. prompt order is fixed as `<|VIDEO|><|AUDIO|><|spatial|>`
+2. processor validates the fixed modal order
+3. base processor expands video and audio placeholders
+4. spatial processor expands spatial placeholders
+5. thinker injects spatial embeddings first
+6. original QwenOmni thinker merges video and audio embeddings afterward
+7. custom RoPE assigns positions by scanning the real `video -> audio -> spatial` token runs
+
+Tokenizer/model synchronization:
+1. instantiate `Qwen2_5OmniSpatialProcessor`
+2. processor registers `<|spatial|>` into `tokenizer`
+3. call `processor.sync_spatial_tokenizer_with_model(model)`
+4. model updates `spatial_token_index` and resizes thinker embeddings
+
+## Verification
+- `python -m py_compile` passed for all newly added scaffold files.
+- A local runtime shape test passed for the new spatial modules without
+  depending on `transformers`:
+  - `features`: `(3, 7, 1000, 64)`
+  - `hidden`: `(3, 200, 128)`
+  - `spatial`: `(3, 47, 256)`
+- Full Qwen integration should still be validated in a target environment that
+  has `transformers`; see
+  [seld233_spatial_validation.md](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/spur-qwen-2.5-omni/docs/seld233_spatial_validation.md).
+
+## What Is Still Intentionally Missing
+- image + spatial joint RoPE path
+- talker-side audio output path when spatial modality is present
+- `use_audio_in_video=True`

docs/seld233_validation.md

@@ -0,0 +1,56 @@
+# SELD233 Spatial Validation
+
+## Purpose
+Validate the implemented `audio + spatial + text` path with a synthetic `B=3`
+FOA batch whose waveform lengths differ per sample.
+
+The validation script prints:
+- processor outputs and placeholder expansion results
+- `FOA -> 7ch` feature shapes
+- SELD backbone hidden-state shapes
+- `2.5 Hz` spatial-token shapes
+- the final multimodal token tensor shape implied by the expanded prompt
+
+If the full Qwen model is loaded, it also builds the actual multimodal
+`inputs_embeds` tensor after audio and spatial injection.
+
+## Script
+- [validate_seld233_spatial_random_foa.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/spur-qwen-2.5-omni/scripts/validate_seld233_spatial_random_foa.py)
+
+## Recommended Run
+Shape-only validation without loading the full Qwen model:
+
+```bash
+python scripts/validate_seld233_spatial_random_foa.py \
+  --model-id /apdcephfs_cq10/share_1603164/user/schmittzhu/model/Qwen2.5-Omni-7B \
+  --baseline-repo-path /apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline \
+  --seld233-checkpoint-path /apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline/3_1_dev_split0_multiaccdoa_foa_model.h5 \
+  --seld233-feature-stats-dir /apdcephfs_cq10/share_1603164/user/schmittzhu/data/seld_feat_label/starss23_plus_foa_16k_29cls
+```
+
+Full multimodal embedding construction on a single device:
+
+```bash
+python scripts/validate_seld233_spatial_random_foa.py \
+  --model-id /apdcephfs_cq10/share_1603164/user/schmittzhu/model/Qwen2.5-Omni-7B \
+  --baseline-repo-path /apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline \
+  --seld233-checkpoint-path /apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline/3_1_dev_split0_multiaccdoa_foa_model.h5 \
+  --seld233-feature-stats-dir /apdcephfs_cq10/share_1603164/user/schmittzhu/data/seld_feat_label/starss23_plus_foa_16k_29cls \
+  --load-qwen-model \
+  --device cuda:0 \
+  --dtype bfloat16
+```
+
+## Expected Core Shapes
+For 20-second padded inputs:
+- `spatial_audio`: `[B, 320000, 4]`
+- `features`: `[B, 7, 1000, 64]`
+- `hidden_states`: `[B, 200, 128]`
+- `spatial_tokens`: `[B, T_spat_max, 256]`
+
+`T_spat_max` depends on the longest valid sample in the batch.
+
+## Current Limits
+- `use_audio_in_video=False` only
+- `image + spatial` not implemented
+- `return_audio=True` not implemented for the spatial subclass

docs/spatial_beats_design.md

@@ -0,0 +1,511 @@
+# Spatial-BEATs 接入 SPUR Spatial QA 的设计说明
+
+## 1. 目标
+
+本文档整理如何将 `Spatial-BEATs` 接入当前的 SPUR/Qwen spatial QA 训练链路，替换现有 `SELD233 + spatial_adapter + spatial_projector` 的空间编码路径。
+
+目标不是修改代码，而是先明确一套可落地的设计：
+
+- 取 `Spatial-BEATs` 的哪一层作为 LLM 的空间 token 源
+- projector/bridge 应该如何设计
+- 如何映射到当前 [train_spur_spatial_qa.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline/train_spur_spatial_qa.py) 的训练模式
+- 如何分阶段训练，尽量复用现有缓存、LoRA、resume、QA 数据集逻辑
+
+---
+
+## 2. 当前 `train_spur_spatial_qa.py` 的关键逻辑
+
+### 2.1 文本接口不是结构化 head，而是 spatial token 占位符
+
+当前 collator 在 prompt 前缀里写入：
+
+```text
+<audio><spatial>
+question...
+```
+
+对应逻辑见：
+
+- [train_spur_spatial_qa.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline/train_spur_spatial_qa.py#L565)
+
+也就是说，LLM 侧只知道有一串 `spatial_token` 要插入，并不要求这串 token 来自某个特定 encoder。
+
+### 2.2 当前空间分支的三层结构
+
+当前 Qwen spatial 路径可以抽象成：
+
+1. `seld233_backbone`
+2. `seld233_spatial_adapter`
+3. `seld233_spatial_projector`
+
+训练模式围绕这三块展开：
+
+- `spatial_only`: 只训练 adapter + projector
+- `adapter_lora`: 冻结空间 backbone，只训练 adapter + projector + LLM LoRA
+- `spatial_lora`: 训练空间 backbone + adapter + projector + LLM LoRA
+
+对应逻辑见：
+
+- [train_spur_spatial_qa.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline/train_spur_spatial_qa.py#L1005)
+- [train_spur_spatial_qa.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline/train_spur_spatial_qa.py#L1059)
+- [train_spur_spatial_qa.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline/train_spur_spatial_qa.py#L1078)
+- [train_spur_spatial_qa.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline/train_spur_spatial_qa.py#L1714)
+
+### 2.3 当前缓存机制的抽象非常适合替换 encoder
+
+collator 目前支持两种缓存输入：
+
+1. feature cache
+   - `seld233_features`: `[B, C, T_feat, M]`
+   - 仍在线运行 backbone
+
+2. hidden cache
+   - `seld233_hidden_states`: `[B, T_hidden, D]`
+   - 直接绕过 feature bridge + backbone
+
+对应逻辑见：
+
+- [train_spur_spatial_qa.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline/train_spur_spatial_qa.py#L630)
+- [train_spur_spatial_qa.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline/train_spur_spatial_qa.py#L672)
+- [train_spur_spatial_qa.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline/train_spur_spatial_qa.py#L1652)
+
+这个抽象意味着：
+
+- 如果 `Spatial-BEATs` 能输出 `[T_s, D]` 的时序 hidden states
+- 那它可以直接替代现有 hidden cache 路径
+
+这对工程落地非常有利。
+
+---
+
+## 3. `Spatial-BEATs` 当前最适合接 LLM 的层
+
+### 3.1 不应直接用监督 head 输出
+
+`Spatial-BEATs` 里存在多种监督读出：
+
+- `slot_latents`
+- `prediction_output`
+- `mono_task_tokens`
+- `mono_prediction_output`
+- `pretrunk_task_tokens`
+- `pretrunk_prediction_output`
+
+这些都不适合直接喂给 LLM，原因是：
+
+- 它们已经是任务特化后的 readout
+- 信息被压缩进分类/方向/距离头
+- 对 QA 这种开放式下游推理来说，保留的上下文冗余不足
+
+### 3.2 推荐层：`fused_spatial_embeddings`
+
+在 `readout_scheme='local_spatial'` 下，最推荐的 LLM 输入源是：
+
+- `fused_spatial_embeddings`
+
+对应逻辑：
+
+- [spatial_beats.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_beats.py#L845)
+- [spatial_beats.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_beats.py#L1092)
+
+它的来源是：
+
+1. `W` 通道经过原始 BEATs semantic path
+2. `WXYZ + IVxyz` 经过 `local_spatial_encoder`
+3. 两路在同一 token rate 上对齐
+4. 经 `local_spatial_proj` 后加和融合
+
+因此这层同时包含：
+
+- 语义稳定性
+- 局部空间 cues
+- 时间结构
+
+这是比单独 `spatial_embeddings` 更好的 LLM 输入层。
+
+### 3.3 备选层：`spatial_embeddings`
+
+如果初版想先降低复杂度，可以先接：
+
+- `spatial_embeddings`
+
+对应逻辑：
+
+- [spatial_beats.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_beats.py#L736)
+- [spatial_beats.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_beats.py#L985)
+
+但在 `local_spatial` 路径下，它少了显式 local-spatial fusion，优先级低于 `fused_spatial_embeddings`。
+
+### 3.4 token 数量建议保持低频时序接口
+
+`Spatial-BEATs` 文档已经明确：
+
+- 最终给 LLM 的应该是低频时序 token
+- 不是 `time x slot`
+- 默认接口为 `2.5 Hz`
+
+对应说明：
+
+- [spatial_beats_token_interface_note.md](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/docs/spatial_beats_token_interface_note.md#L32)
+
+因此推荐的外部接口是：
+
+- `10s -> 25 tokens`
+- 形状：`[B, T_s, D]`
+
+不要把内部 slot 数 `K` 展开成 `T_s * K` 个 LLM token。
+
+---
+
+## 4. 推荐的 bridge/projector 设计
+
+### 4.1 当前 `Spatial-BEATs` 自带 projector
+
+当前 projector 为：
+
+- `Linear(D -> H)`
+- `GELU`
+- `LayerNorm(H)`
+- `Linear(H -> d_llm)`
+
+见：
+
+- [spatial_modules.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/spatial_modules.py#L1159)
+
+这作为第一版是合理的，但它是“纯 token-wise MLP”，没有显式时间建模。
+
+### 4.2 推荐：LLM bridge 版 projector
+
+针对 QA 下游，推荐的 bridge 设计不是只做线性升维，而是：
+
+```text
+fused_spatial_embeddings [B, T_s, D]
+  -> LayerNorm
+  -> temporal bridge block
+  -> token projector MLP
+  -> llm_spatial_tokens [B, T_s, d_llm]
+```
+
+其中：
+
+#### 1. temporal bridge block
+
+建议用轻量模块，1 层即可：
+
+- 1-layer TransformerEncoder
+或
+- gated MLP / temporal Conv1d block
+
+目的：
+
+- 在进入 LLM 前，对低频空间 token 再做一次跨时间整形
+- 补偿 projector 完全逐 token 独立的缺陷
+
+推荐输入输出维度不变：
+
+- `D = 768`
+
+#### 2. token projector MLP
+
+建议结构：
+
+```text
+LayerNorm(768)
+Linear(768, 1536)
+GELU
+Linear(1536, 4096)
+```
+
+如果想更稳，可以再加：
+
+- `LayerNorm(4096)`
+或
+- 一个可学习 gate
+
+例如：
+
+```text
+proj = MLP(x)
+gate = sigmoid(W_g x)
+llm_tokens = gate * proj
+```
+
+但第一版不必上 gate。
+
+### 4.3 为什么不建议直接投分类/角度/距离结果
+
+不建议将这些 head 输出拼成 LLM 输入：
+
+- `pred_class_logits`
+- `pred_direction`
+- `pred_distance`
+
+原因：
+
+- 这些已经是强任务约束后的压缩表示
+- 会丢掉很多时序上下文和不确定性信息
+- LLM 更适合消费 dense latent，而不是 encoder 的最终判决
+
+这些输出可以作为：
+
+- 训练时的 auxiliary supervision
+- 或后续做解释性 probing
+
+但不应作为主 token 接口。
+
+---
+
+## 5. 与当前 `train_spur_spatial_qa.py` 的映射设计
+
+### 5.1 模块映射
+
+建议把当前 `SELD233` 路径替换为：
+
+```text
+seld233_backbone          -> spatial_beats_backbone
+seld233_spatial_adapter   -> spatial_beats_bridge
+seld233_spatial_projector -> spatial_beats_llm_projector
+```
+
+更具体地说：
+
+#### `spatial_beats_backbone`
+
+包含：
+
+- preprocessor
+- BEATs trunk
+- local_spatial_encoder
+- local_spatial_proj
+- local_spatial_fusion_norm
+
+输出：
+
+- `fused_spatial_embeddings`
+- `temporal_padding_mask`
+
+#### `spatial_beats_bridge`
+
+新增的一层轻量 temporal bridge：
+
+- 1-layer temporal adapter
+- 输入输出都为 `768`
+
+这是新设计里最关键的“LLM bridge”部分。
+
+#### `spatial_beats_llm_projector`
+
+将 bridge 后的 token 投到 LLM hidden size：
+
+- `768 -> 1536 -> 4096`
+
+输出：
+
+- `[B, T_s, d_llm]`
+
+### 5.2 对应当前两种 cache 语义
+
+推荐保留当前 cache 设计思路，但改名语义如下：
+
+#### feature cache
+
+缓存 `Spatial-BEATs` 前端中间结果，例如：
+
+- `foa_feat`
+- 或 patch/token 级中间特征
+
+只在你想继续训练 backbone 时使用。
+
+#### hidden cache
+
+缓存：
+
+- `fused_spatial_embeddings`
+或
+- `bridge` 之前的 `[T_s, 768]`
+
+这是最有用的 cache 形式。
+
+它对应当前 `train_spur_spatial_qa.py` 的 hidden-cache 路线：
+
+- 在线不跑空间 backbone
+- 只训练 bridge/projector/LLM
+
+对于 QA 阶段，这会是最实用的加速方案。
+
+---
+
+## 6. 推荐训练阶段
+
+### 阶段 A：encoder-only 预训练
+
+目标：
+
+- 在 `Spatial-BEATs` 自己的监督任务上训出靠谱的 `fused_spatial_embeddings`
+
+推荐 checkpoint：
+
+- `ov1_local_spatial` 最优 checkpoint
+- 更理想是其 balanced follow-up checkpoint
+
+此阶段不接 LLM。
+
+### 阶段 B：freeze Spatial-BEATs，训练 bridge + LLM LoRA
+
+这是最推荐的第一轮接入实验。
+
+训练参数对应当前 QA 框架可映射为：
+
+- 冻结 `spatial_beats_backbone`
+- 训练 `spatial_beats_bridge`
+- 训练 `spatial_beats_llm_projector`
+- 训练 `LLM LoRA`
+
+它在当前 `train_spur_spatial_qa.py` 中最接近：
+
+- `--train-adapter-lora`
+
+只是把现有的 `seld233_spatial_adapter + seld233_spatial_projector`
+替换成 `spatial_beats_bridge + spatial_beats_llm_projector`。
+
+这是最值得先做的实验，因为：
+
+- encoder 先不动，稳定
+- 速度快
+- 容易判断 `Spatial-BEATs` token 是否比 DCASE encoder 更有用
+
+### 阶段 C：解冻 Spatial-BEATs 尾部，小学习率联合微调
+
+当阶段 B 证明 token 有效后，再解冻少量模块：
+
+- `local_spatial_proj`
+- `local_spatial_fusion_norm`
+- 可选 `local_spatial_encoder` 最后一层
+- 可选 BEATs trunk 最后 1-2 层
+
+训练对象：
+
+- Spatial-BEATs 尾部
+- bridge
+- projector
+- LLM LoRA
+
+它在当前 QA 框架中最接近：
+
+- `--train-spatial-lora`
+
+但不建议一开始就这样做，因为会把“encoder 不好”与“LLM 对不上”混在一起。
+
+### 阶段 D：hidden-cache 加速 QA 训练
+
+当阶段 B/C 结构稳定后，推荐预先导出：
+
+- `fused_spatial_embeddings`
+- `temporal_padding_mask`
+
+作为 hidden cache。
+
+然后 QA 训练使用 hidden cache，避免每次在线跑 Spatial-BEATs。
+
+这对应当前脚本里的 hidden cache 模式：
+
+- [train_spur_spatial_qa.py](/apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline/train_spur_spatial_qa.py#L630)
+
+---
+
+## 7. 推荐的第一版落地方案
+
+如果只做第一版、最小风险接入，建议如下：
+
+### 7.1 encoder 输出
+
+从 `Spatial-BEATsOutput` 里读取：
+
+- 首选 `fused_spatial_embeddings`
+- fallback `spatial_embeddings`
+
+同时读取：
+
+- `temporal_padding_mask`
+
+### 7.2 bridge/projector
+
+使用：
+
+```text
+fused_spatial_embeddings [B, T_s, 768]
+  -> LayerNorm
+  -> 1-layer TransformerEncoder (768)
+  -> LayerNorm
+  -> Linear(768, 1536)
+  -> GELU
+  -> Linear(1536, 4096)
+```
+
+输出：
+
+- `llm_spatial_tokens [B, T_s, 4096]`
+
+### 7.3 QA 训练策略
+
+第一轮只做：
+
+- freeze Spatial-BEATs backbone
+- train bridge
+- train projector
+- train LLM LoRA
+
+也就是概念上对应当前的：
+
+- `adapter_lora`
+
+### 7.4 不建议第一版就做的事
+
+以下都不建议第一版就做：
+
+- 用 slot 输出替代 dense latent
+- 同时训练 Spatial-BEATs 全 backbone
+- 同时改 QA prompt 设计
+- 把分类/角度/距离 head 输出直接拼进 LLM token
+
+这些会大幅增加排障复杂度。
+
+---
+
+## 8. 为什么这条路线比当前 DCASE encoder 更合理
+
+当前 DCASE spatial encoder 路径的问题是：
+
+- 更偏任务 readout
+- 对 QA 来说 token 表征容量不足
+- 数值空间信息学得不稳
+
+而 `Spatial-BEATs local_spatial` 路径的优势是：
+
+1. 有明确的低频时序 token 接口
+2. `fused_spatial_embeddings` 本身就是为“投给 LLM”准备的 dense latent
+3. semantic path 与 local spatial path 分工清晰
+4. 已经有 projector 抽象和 token interface 文档
+
+换句话说，`Spatial-BEATs` 的结构天然更像一个可插拔的独立 spatial encoder，而不是一个“SELD head 顺便挤出 token”。
+
+---
+
+## 9. 建议的下一步
+
+建议按下面顺序推进：
+
+1. 先实现 `Spatial-BEATs -> fused_spatial_embeddings -> bridge -> projector -> LLM spatial tokens`
+2. 第一轮只做 freeze encoder + bridge/projector + LLM LoRA
+3. 验证 QA 上是否比当前 DCASE spatial encoder 明显更好
+4. 如果有效，再做尾部联合微调
+5. 最后再做 hidden-cache 预计算优化
+
+这样可以把风险拆开：
+
+- encoder token 是否有效
+- bridge 是否够用
+- LLM 是否能消费这组 token
+
+而不是把三个问题同时耦合起来。
+

docs/spatial_encoder_plan.md

@@ -0,0 +1,417 @@
+# SELDNet-233 Spatial Modality Integration Design
+  ## Document Status
+  - Owner: Student Implementation
+  - Reviewer: Project Maintainer
+  - Target Repository: `/apdcephfs_cq10/share_1603164/user/schmittzhu/code/spur-qwen-2.5-omni`
+  - Related Baseline Repository: `/apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline`
+
+  ## 1. Background
+  This project aims to extend the current `Qwen2.5-Omni`-based system with a new **independent spatial modality** derived from a pretrained `SELDNet` encoder.
+  The current system already supports:
+  - text
+  - image
+  - video
+  The current SPUR-related spatial path in the Omni codebase is not a truly independent modality. Instead, it computes spatial features and fuses them back into the audio encoder branch. This is
+  **not** the desired design for this project.
+
+  - keep the **original audio encoder unchanged**
+  - use the same FOA input audio in **two parallel paths**
+  - inject a new set of **spatial tokens** into the LLM as a separate modality
+  ## 2. Objective
+  Given a 4-channel FOA audio input:
+  1. The **W channel** should continue to go through the original Omni mono audio pipeline.
+  2. The full **4-channel FOA** should go through a pretrained `SELDNet-233` encoder.
+  3. The `SELDNet-233` encoder should provide a sequence of **spatial tokens** derived from its **MHSA output**, not from its final localization/classification heads.
+  4. These spatial tokens should be projected into the LLM hidden space and injected into the thinker decoder input as a new modality `<|spatial|>`.
+
+  ## 3. Non-Goals
+  - modify the existing SELD baseline source code
+  - modify existing SELD training or inference behavior
+  - use SELDNet final `accdoa` or `sed_logits` as the LLM input representation
+
+  ## 4. Fixed Dependencies
+  - Path: `/apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline`
+
+  ### 4.2 SELD checkpoint to use
+
+  ### 4.3 SELD task configuration
+  - Task ID: `233`
+
+  ### 4.4 SELD model definition
+
+  ## 5. Key Design Decision
+
+  ### 5.1 Spatial representation source
+  The spatial representation must come from the **shared temporal representation after the final MHSA block** in `SELDNet`, not from:
+  - final `accdoa`
+  - final `sed_logits`
+  - explicit detection heads
+
+  ### 5.2 Why this representation
+  This tensor is:
+  - temporally ordered
+  - dense
+  - scene-level
+  - richer than detection output heads
+  - better suited for LLM conditioning
+
+  ## 6. Target Architecture
+
+  ### 6.1 Input
+
+  ### 6.2 Dual-path processing
+  - take FOA `W channel`
+  - convert to mono
+  - feed into original Omni audio encoder
+  - keep existing tokenization rate unchanged
+  - expected token rate: about `25 Hz`
+  Path B: New spatial path
+  - keep all 4 FOA channels
+  - convert to SELD baseline-compatible input features
+  - feed into pretrained `SELDNet-233`
+  - produce spatial token sequence
+  - project into LLM hidden size
+  - inject as `<|spatial|>` modality
+  ### 6.3 Final LLM view
+  - normal text tokens
+  - normal audio tokens from original audio encoder
+  - optional image/video tokens
+  - new spatial tokens from SELDNet-233 branch
+
+  ## 7. Temporal Resolution Requirements
+
+  ### 7.1 Audio token frequency
+  Keep current Omni audio token frequency unchanged:
+  - approximately `25 Hz`
+
+  ### 7.2 Spatial token frequency
+  Spatial tokens should be lower frequency:
+  - target `2.5 Hz`
+
+  This is intentional because:
+  - lower token rate reduces prompt length
+  - lower token rate is sufficient for spatial reasoning
+  Recommended first version:
+  - `h_seld`: `[B, T_seld, 128]`
+  - `spatial_tokens`: `[B, T_spat, 256]`
+  - `projected_spatial`: `[B, T_spat, D_llm]`
+
+  Where:
+  - `T_spat = 5 * seconds`
+
+  ## 8. Data Flow
+
+  1. Input FOA waveform enters processor.
+  2. Processor keeps original audio path intact:
+     - use `W channel` for original audio encoder.
+  3. Processor also passes full FOA waveform as `spatial_audio`.
+     - `spatial_audio` for new spatial path
+  5. New SELD spatial adapter computes:
+     - SELD MHSA output
+     - `2.5 Hz` spatial tokens
+  6. Spatial projector maps tokens to LLM hidden dimension.
+  8. Thinker injects projected spatial embeddings into `inputs_embeds`.
+  9. LLM decoder consumes them as a true multimodal prefix.
+
+  ## 9. Required Code Work
+
+
+  ### A. `qwen2_5_omni_spur/modules/seldnet233_backbone.py`
+  Purpose:
+  - build baseline `SeldModel` with task `233`
+  - load pretrained checkpoint
+  - freeze model by default
+  - register a forward hook on the final MHSA output
+
+  Expected behavior:
+  - input: baseline-compatible SELD features
+  - output: `h_seld [B, T_seld, D_seld]`
+
+  - do not modify baseline code
+  - use import-time wrapping or dynamic module loading
+  - use forward hook on the last MHSA normalization output
+
+  ### B. `qwen2_5_omni_spur/modules/seldnet233_feature_bridge.py`
+  Purpose:
+  - convert FOA waveform into the feature representation expected by SELD task `233`
+
+  Expected behavior:
+  - output: baseline feature tensor `[B, C, T_feat, F_feat]`
+
+  Implementation notes:
+  - must match SELD task `233` feature config exactly
+  - sample rate must be `16k`
+  - should reuse baseline feature logic where possible
+
+  Purpose:
+  - build a full SELD-based spatial token extractor
+
+  Internal structure:
+  - feature bridge
+  - SELD backbone
+  - spatial token head
+
+  Expected behavior:
+  - output:
+    - `spatial_tokens [B, T_spat, D_spat]`
+
+  Implementation notes:
+  - token rate target must be `2.5 Hz`
+  - default token dim should be `256`
+  - token head may use:
+    - pooling
+    - MLP
+  - recommended first implementation:
+
+  ### D. Optional: `qwen2_5_omni_spur/modules/spatial_token_projector.py`
+  Purpose:
+  - project SELD spatial tokens into LLM hidden dimension
+
+
+  ## 9.2 Existing files to modify
+  ### A. `qwen2_5_omni_spur/configuration_qwen2_5_omni.py`
+  Add config fields:
+  - `spatial_token_index`
+  - `use_seld233_spatial_modality`
+  - `seld233_checkpoint_path`
+  - `seld233_token_dim`
+  - `seld233_token_rate_hz`
+  - `seld233_projector_hidden_dim`
+  - optional `seld233_freeze_backbone`
+
+  - `attribute_map` to include `spatial_token_id`
+
+  ### B. `qwen2_5_omni_spur/modules/__init__.py`
+  Export new modules:
+  - `SeldNet233Backbone`
+  - optional new projector class
+
+  ### C. `qwen2_5_omni_spur/processing_qwen2_5_omni.py`
+  Add a new special token:
+  - `<|spatial|>`
+
+  - store `self.spatial_token`
+  - optionally accept `spatial_tokens` and `spatial_token_lengths`
+  - expand `<|spatial|>` into repeated placeholders based on spatial token length
+  - expose spatial fields through `model_input_names`
+  Important constraint:
+  - original W-channel mono audio path must remain unchanged
+  ### D. `qwen2_5_omni_spur/modeling_qwen2_5_omni.py`
+  Required changes:
+  - instantiate SELD spatial adapter in thinker
+  - accept `spatial_audio` in thinker forward
+  - compute spatial embeddings
+  - inject them via `masked_scatter`
+  - update multimodal rope logic
+  - route spatial fields through generation
+  ### E. `training-qwen-omni/train_spur_spatial_only_hf.py`
+  Update collator:
+  - additionally pass full FOA as `spatial_audio`
+
+  ### F. `qwen-omni-inference.py`
+  Update inference script:
+  - register `<|spatial|>` token
+  - construct prompt with spatial placeholder
+  - validate both paths run together
+
+  ## 10. Processor Design
+
+  ### 10.1 Prompt convention
+  - `<|spatial|>`
+
+  Example:
+  - `<|AUDIO|><|spatial|> Please answer the question using both sound content and spatial cues.`
+
+  ### 10.2 Placeholder expansion behavior
+  - `T_spat` repeated spatial token positions
+
+  This mirrors current audio/image/video placeholder logic.
+
+  ### 10.3 Recommended input mode
+  - processor receives `spatial_audio`
+
+  Fallback mode:
+  - processor receives precomputed `spatial_tokens`
+
+  ## 11. SELD Path Design
+  ### 11.1 Backbone loading
+  The SELD branch must:
+  - use task `233`
+  - load the fixed checkpoint above
+  - stay frozen in the first training stage
+  ### 11.2 Feature extraction
+  The spatial branch must operate on:
+  - full 4-channel FOA
+  - `16kHz`
+  - baseline-compatible feature format
+  ### 11.3 Hidden representation extraction
+  The representation to extract is:
+  - final MHSA output
+  - not final heads
+  - not sed logits
+
+  ### 11.4 Tokenization
+  The spatial token head must:
+  - preserve temporal ordering
+  - reduce token rate to `2.5 Hz`
+  - output a fixed hidden dim, recommended `256`
+
+  ### 12.1 New thinker members
+  Add:
+  - optional `self.seld233_spatial_norm`
+  ### 12.2 Forward inputs
+  Thinker forward should support:
+  - `spatial_audio`
+  - `spatial_tokens`
+  - `spatial_token_lengths`
+
+  Priority:
+  - if `spatial_tokens` provided, use them directly
+
+  ### 12.3 Injection point
+  Injection should happen in the same stage where current code merges:
+  - text
+  - audio
+  - video
+  The new logic should:
+  1. compute `projected_spatial`
+  2. build `spatial_mask = (input_ids == spatial_token_id)`
+  3. inject with `masked_scatter`
+
+  Before scatter, verify:
+
+  If not, raise an error.
+
+
+  ### 13.1 Why needed
+  Adding a new modality requires updating multimodal position ID construction.
+  ### 13.2 Required additions
+  - add `spatial_token_id`
+  - add `spatial_seqlens`
+  - count spatial modality occurrences
+  - add `remain_spatials`
+  - add spatial branch when selecting next multimodal segment
+
+  ### 13.3 Spatial position assignment
+  Spatial token positions should be assigned:
+  - length equal to `spatial_token_lengths`
+  - consistent with `2.5 Hz` token rate
+
+  ## 14. Generation Path Changes
+
+  Top-level `generate()` must forward:
+  - `spatial_audio`
+  - `spatial_tokens`
+  - `spatial_token_lengths`
+  to thinker generation.
+  ### 14.2 Talker interaction
+  Recommended first version:
+  - spatial positions should be zeroed before passing thinker hidden states into talker
+
+  Reason:
+  - spatial modality is meant for reasoning
+  - not necessarily for speech token generation conditioning
+  ## 15. Training Strategy
+
+  ### 15.1 Stage 1
+  Freeze:
+  - SELD backbone
+
+  Train only:
+  - spatial token head
+  - spatial projector
+  - LoRA / lightweight LLM adapters if used
+
+  - last spatial token head layers
+  - projector
+  - maybe a small part of SELD adapter
+  ### 15.3 Stage 3
+  Optional advanced finetuning:
+  - limited SELD backbone finetuning
+  - only after full pipeline is stable
+  ## 16. Validation Plan
+  ### 16.1 Unit-level validation
+  - checkpoint loads successfully
+  - MHSA output is captured successfully
+  - spatial token head outputs expected shape
+  - token rate is `2.5 Hz`
+  ### 16.2 Processor validation
+  Validate:
+  - spatial fields appear in batch output
+  - original audio path is unchanged
+
+  Validate:
+  - spatial projector output matches LLM hidden size
+  - no length mismatch
+  - no dtype/device mismatch
+  ### 16.4 Rope validation
+  Validate:
+  - generation does not crash when `<|spatial|>` is present
+
+  Validate:
+  - original audio tokens still exist
+  - spatial tokens are injected independently
+  - text generation completes
+
+  ## 17. Risks
+
+  ### 17.1 Feature mismatch risk
+  If SELD features differ from training-time `233` config, checkpoint utility may collapse.
+
+  Mitigation:
+  - strictly reuse baseline feature logic
+
+  If hook location is incorrect, extracted representation may not correspond to post-MHSA scene representation.
+  Mitigation:
+  - verify against baseline layer order
+  - inspect shape and stability
+
+  ### 17.3 Rope mismatch
+  If `<|spatial|>` positions are not included in multimodal rope indexing, model behavior may degrade or fail.
+
+  Mitigation:
+  - treat spatial exactly as a fourth modality in rope construction
+
+  ### 17.4 Prompt length growth
+  Adding spatial tokens increases context length.
+  Mitigation:
+
+  ## 18. Final Requirements Checklist
+
+  - [ ] Use checkpoint `233_merged29_foa_16k_sedwarmup_v1_dev_split0_multiaccdoa_foa_best_full_model.h5`
+  - [ ] Keep original audio encoder unchanged
+  - [ ] Audio path uses `W channel` only
+  - [ ] Spatial path uses all 4 FOA channels
+  - [ ] Spatial representation comes from final SELD MHSA output
+  - [ ] Spatial token rate is `2.5 Hz`
+  - [ ] Spatial token dimension defaults to `256`
+  - [ ] Spatial tokens are injected as a true independent modality
+  - [ ] `<|spatial|>` placeholder is supported in processor
+  - [ ] RoPE logic includes spatial modality
+  - [ ] End-to-end generate path runs successfully
+
+  ## 19. Recommended Implementation Order
+
+  1. Implement `seldnet233_backbone.py`
+  2. Implement `seldnet233_feature_bridge.py`
+  3. Implement `seldnet233_spatial_adapter.py`
+  4. Add config fields
+  5. Add processor `<|spatial|>` support
+  6. Add thinker spatial encoder + projector
+  7. Add spatial `masked_scatter`
+  8. Update `get_rope_index()`
+  9. Update `generate()`
+  10. Update collator and inference script
+  11. Run shape-only validation
+  12. Run end-to-end generation validation
+
+  ## 20. Reviewer Notes
+
+  This design intentionally separates:
+  - content understanding from the original audio encoder
+  - spatial reasoning from the SELDNet-based encoder
+
+  The design is successful only if the final model sees:
+  - mono audio content tokens at normal resolution
+  - spatial tokens at low resolution
+  - both as independent modalities inside the thinker input stream

docs/spatial_qwen3_design.md

@@ -0,0 +1,821 @@
+# Spatial-Qwen 升级到 Qwen3-Omni-MoE-30B-A3B 设计文档
+
+> **状态**：阶段 0–4 已完成；阶段 5（端到端 smoke + 1 epoch 真实训练）和阶段 1d（pytest 集成测试）待执行。
+> **最后更新**：2026-05-11
+> **作者**：Claude Code（claudeMd 项目协作）
+>
+> 本文档是 Qwen3-Omni-MoE Spatial 升级的**单一权威记录**。
+> 旧路径（Qwen2.5-Omni-7B Spatial）完全不动；本工程独立运行在新 conda env `spur-qwen3` 上，
+> 使用独立的 model 类、独立的训练入口、独立的 RUN_ROOT。
+
+---
+
+## 0. 总览
+
+把 Spatial-Qwen 的 Spatial-BEATs 路径（FOA 4ch 音频 → spatial token → `<|spatial|>` 占位符 → LLM）
+从 backbone **Qwen2.5-Omni-7B (Dense Thinker, hidden=4096)** 升级到
+**Qwen3-Omni-30B-A3B-Instruct (MoE Thinker, hidden=2048, 128 experts/top-8)**。
+
+### 设计准则（按用户决策固化）
+
+| 决策点 | 选择 | 原因 |
+|---|---|---|
+| Qwen3 模型规格 | **30B-A3B-Instruct** | 用户明确指定；Instruct 版省了 chat template 调优 |
+| 集成模式 | **重侵入式**（subclass + override `forward` / `__init__`） | 与 Qwen2.5 现有 `Qwen2_5OmniSpatialThinkerForConditionalGeneration` 对齐；不走 AF3 轻量 wrapper 模式 |
+| Spatial encoder | **仅 Spatial-BEATs** | IV / Neural-IV / SELD233 留到下一轮；先验证 30B-MoE 路径能跑起来 |
+| Conda env | **克隆 `spur` 为 `spur-qwen3`** | 与现有训练环境完全隔离；`pip install` 不污染 spur |
+| transformers 来源 | **本地 fork editable install** | pip 4.52 没有 qwen3_omni_moe；fork 5.0.0.dev0 有，但 talker / video_processor 有 bug |
+| 现有代码 | **完全不动** | 所有 Qwen2.5 / IV / AF3 / shell / runs / ckpt 保持不变 |
+
+### 硬件 / 软件约束
+
+- **GPU**：8× A100 40 GB
+- **CUDA driver**：12.2
+- **CUDA runtime (torch)**：12.8（torch 2.10）
+- **conda**：conda 25.x，env 安装到 `/opt/conda/envs/spur-qwen3`（12T overlay，8.5T free；apdcephfs_cq10 共享 NFS 已仅剩 3.7 GB 可用）
+
+---
+
+## 1. 阶段 0：环境（已完成）
+
+### 1.1 克隆 conda env
+
+```bash
+conda create -n spur-qwen3 --clone spur
+conda activate spur-qwen3
+```
+
+完成后 `python --version` = Python 3.10.19，与 `spur` 一致。
+
+### 1.2 安装本地 transformers fork (editable)
+
+`pip install transformers==4.52` 没有 `qwen3_omni_moe` 模块。本地已有完整 fork：
+
+```
+/apdcephfs_cq10/share_1603164/user/schmittzhu/model/transformers/src/transformers/models/qwen3_omni_moe/
+    configuration_qwen3_omni_moe.py     # Thinker / Talker / Audio / Vision config
+    modeling_qwen3_omni_moe.py          # 4138 行，Thinker 在 line 1906
+    modular_qwen3_omni_moe.py           # 模块化源码（HF tooling 自动生成 modeling）
+    processing_qwen3_omni_moe.py        # Qwen3OmniMoeProcessor + ProcessorKwargs
+```
+
+实际安装命令（以 fork 根目录的 `pyproject.toml` 为入口）：
+
+```bash
+cd /apdcephfs_cq10/share_1603164/user/schmittzhu/model/transformers
+pip install -e .
+```
+
+安装后 `transformers.__version__` = `5.0.0.dev0`（editable）。
+
+### 1.3 修复 fork 引入的依赖冲突
+
+| 现象 | 原因 | 修复 |
+|---|---|---|
+| `huggingface_hub.is_offline_mode` 不存在 | hf_hub 0.36.2 与 transformers 5.0 不兼容 | `pip install -U "huggingface-hub>=1.3.0,<2.0"` → 1.14.0 |
+| `tokenizers` 兼容性报错 | tokenizers 0.21.4 旧 | `pip install -U "tokenizers>=0.22.0,<=0.23.0"` → 0.22.2 |
+
+### 1.4 拍 fork 快照（防止改坏后无法回滚）
+
+```bash
+cp -r /apdcephfs_cq10/share_1603164/user/schmittzhu/model/transformers \
+      /apdcephfs_cq10/share_1603164/user/schmittzhu/model/transformers.snapshot-pre-qwen3
+```
+
+### 1.5 已知 fork bug 与规避
+
+1. **`Qwen3OmniMoeConfig.from_pretrained(MODEL)` 顶层加载崩溃**
+   - 错误：`AttributeError: 'Qwen3OmniMoeTalkerCodePredictorConfig' object has no attribute 'use_sliding_window'`
+   - 规避：直接读 `config.json` 原始 JSON，从 `raw["thinker_config"]` 实例化 `Qwen3OmniMoeThinkerConfig`，**完全跳过 talker 解析**。我们不需要 talker，所以这是无成本规避。
+2. **`AutoProcessor.from_pretrained` 崩溃**
+   - 错误：`TypeError: argument of type 'NoneType' is not iterable` 在 video_processor auto-loading
+   - 规避：分别加载 `AutoTokenizer` + `AutoFeatureExtractor`，跳过 image / video processor。
+3. **`AutoVideoProcessor` 需要 torchvision**，spur-qwen3 env 里没装；不打算装（音频 QA 不用 video）。
+4. **`Qwen3OmniMoeProcessor.__init__` 拒绝 None image / video processor**
+   - 内部 `check_argument_for_proper_class(image_processor, ImageProcessingMixin)` 抛 `TypeError`
+   - 规避：subclass 里**不调用** `super().__init__`，手动赋值 attribute（见 §3.2）
+
+### 1.6 pip freeze（关键依赖）
+
+```
+python              : 3.10.19
+torch               : 2.10.0+cu128 (CUDA 12.8 runtime；driver 12.2 兼容)
+torchaudio          : 2.10.0
+torchvision         : NOT installed（不需要）
+transformers        : 5.0.0.dev0  (editable @ /apdcephfs_cq10/.../model/transformers)
+huggingface_hub     : 1.14.0
+tokenizers          : 0.22.2
+peft                : 0.18.1
+deepspeed           : 0.19.0  （已装，但 train script 暂未 wire；见 §6）
+flash_attn          : NOT installed → ATTN_IMPL=sdpa
+soundfile / numpy / tqdm / einops / safetensors / accelerate
+```
+
+### 1.7 激活 env
+
+```bash
+source /opt/conda/etc/profile.d/conda.sh
+conda activate spur-qwen3
+```
+
+> 现有 Qwen2.5 / IV / AF3 训练继续在 `spur` env 跑，**不受影响**。
+
+### 1.8 Qwen3 模型权重
+
+```
+/apdcephfs_cq12/share_302080740/model/Qwen3-Omni-30B-A3B-Instruct/
+    config.json                # 含 thinker_config + talker_config + token2wav_config
+    model-{00001..00015}-of-00015.safetensors   # 30B BF16 ≈ 60 GB，分 15 个分片
+    tokenizer.json / vocab.json / merges.txt
+    chat_template.jinja / preprocessor_config.json
+```
+
+---
+
+## 2. 阶段 1：spatial 子类骨架（已完成）
+
+### 2.1 Qwen3 Thinker config 关键参数（实测）
+
+```
+hidden_size           = 2048      （Qwen2.5: 4096） ← projector output_dim 必须改！
+num_hidden_layers     = 48        （Qwen2.5: 28）
+num_experts           = 128       （MoE）
+num_experts_per_tok   = 8
+mrope_section         = (24, 20, 20)   （Qwen2.5: (16, 24, 24)）
+mrope_interleaved     = True            （Qwen2.5: 不 interleaved）
+audio_token_id        = 151675   （Qwen2.5: 151646）
+image_token_id        = 151655
+video_token_id        = 151656
+audio_encoder         = qwen3_omni_moe_audio_encoder（Whisper-flavor，128 mel）
+sampling_rate         = 16000
+audio_token (str)     = "<|audio_pad|>"          ← 关键差异！Qwen2.5 用 "<|AUDIO|>"
+audio_bos / audio_eos = "<|audio_start|>" / "<|audio_end|>"
+```
+
+### 2.2 文件 1：`spatial_qwen/model/configuration_qwen3_omni.py`（202 行）
+
+**目的**：定义 spatial 扩展 config，subclass `Qwen3OmniMoeThinkerConfig`，加 30+ 个 `spatial_*` 字段。
+
+**核心代码**：
+
+```python
+class Qwen3OmniMoeSpatialThinkerConfig(Qwen3OmniMoeThinkerConfig):
+    model_type = "qwen3_omni_moe_thinker_spatial"
+    attribute_map = {"spatial_token_id": "spatial_token_index"}
+    sub_configs = {
+        "audio_config": Qwen3OmniMoeAudioEncoderConfig,
+        "vision_config": Qwen3OmniMoeVisionEncoderConfig,
+        "text_config": Qwen3OmniMoeTextConfig,
+    }
+
+    def __init__(self, ..., spatial_token_index=None, spatial_encoder_type="spatial_beats", ..., **kwargs):
+        kwargs.pop("model_type", None)   # 防止 from_pretrained roundtrip 时父类把 model_type 改回 vanilla
+        # 设置所有 spatial_* 字段
+        self.spatial_token_index = spatial_token_index
+        ...
+        # 把 audio/vision/text + token id 交给父类
+        super().__init__(audio_config=..., vision_config=..., text_config=..., **kwargs)
+```
+
+**关键设计**：
+- `model_type` 用唯一名字 `qwen3_omni_moe_thinker_spatial`，避免 AutoConfig 误匹配。
+- `attribute_map` **只**保留 `spatial_token_id → spatial_token_index`（Qwen3 父类直接用 `*_token_id`，但旧 spatial 代码读 `config.spatial_token_index`，所以保留这一个映射）。
+- `kwargs.pop("model_type", None)`：HF `to_dict()` 把 `model_type` 序列化进 dict，下次 `from_dict(**d)` 父类 `__init__` 会把 `self.model_type` 设回 `qwen3_omni_moe_thinker`，覆盖类级属性。必须先 pop 才能保住 subclass 的 `model_type`。
+- spatial 字段顺序：先 BEATs（10 个），再 IV（9 个），再 SELD233（16 个），与 Qwen2.5 `configuration.py` lines 583–676 完全对齐。
+- `spatial_encoder_type` 默认值改为 `"spatial_beats"`（Qwen2.5 默认 `"seld233"`）。
+- 添加 `llm_hidden_size` property（=`text_config.hidden_size`，Qwen3 上是 2048）。
+
+**Smoke 测试**（基于真实 30B config.json）：
+
+```python
+import json
+from spatial_qwen.model.configuration_qwen3_omni import Qwen3OmniMoeSpatialThinkerConfig
+raw = json.load(open('/apdcephfs_cq12/.../Qwen3-Omni-30B-A3B-Instruct/config.json'))
+cfg = Qwen3OmniMoeSpatialThinkerConfig(**raw['thinker_config'], spatial_token_index=151700)
+# → hidden=2048, layers=48, audio_token_id=151675, model_type=qwen3_omni_moe_thinker_spatial
+# → roundtrip cfg.to_dict() → Qwen3OmniMoeSpatialThinkerConfig(**d) 字段全部保住
+```
+
+### 2.3 文件 2：`spatial_qwen/model/processing_spatial_qwen3.py`（494 行）
+
+**目的**：Qwen3-Omni-MoE 的 spatial 处理器（仅 BEATs 路径），处理 FOA 输入并构造 `<|spatial|>` 占位符。
+
+**与 Qwen2.5 spatial processor 的关键差异**：
+
+1. **不走 SELD233 / IV 显式 payload 分支**（删掉了 ~200 行）。
+2. **Qwen3 `audio_token` 字符串是 `<|audio_pad|>`**，不是 `<|AUDIO|>`。`audio_token_aliases` 包含 `("<|audio|>", "<|AUDIO|>")` 两种，方便用户写习惯的 prompt。
+3. **不调 `super().__init__`**（fork 父类拒绝 None image/video processor）：
+   ```python
+   def __init__(self, image_processor=None, video_processor=None, feature_extractor=None,
+                tokenizer=None, chat_template=None):
+       # 手动赋 attribute 跳过 ProcessorMixin 的类型检查
+       self.image_processor = image_processor       # None
+       self.video_processor = video_processor       # None
+       self.feature_extractor = feature_extractor
+       self.tokenizer = tokenizer
+       self.chat_template = chat_template
+       # 直接读 tokenizer 的 special token 字符串
+       self.audio_token = self.tokenizer.audio_token   # "<|audio_pad|>"
+       self.image_token = getattr(self.tokenizer, "image_token", "<|IMAGE|>")
+       ...
+       self.attributes = []  # ProcessorMixin enumerate 子组件用，留空
+       self.spatial_token = "<|spatial|>"
+       self.spatial_token_id = self._register_spatial_token_on_tokenizer()
+   ```
+4. **`__call__` 不调 `super().__call__`**（父类会触碰 image/video 分支，但 torchvision 没装）。改为内联实现 audio-only 路径：
+   ```python
+   audio_inputs = self.feature_extractor(upstream_audio, sampling_rate=16000, padding=True, ...)
+   audio_inputs["feature_attention_mask"] = audio_inputs.pop("attention_mask")
+   audio_lengths = _get_feat_extract_output_lengths(feat_attn.sum(-1))
+   # expand <|AUDIO|> 占位符为 per-frame audio_pad token
+   for sample in text_list:
+       count = sample.count(self.audio_token)
+       for _ in range(count):
+           sample = sample.replace(self.audio_token, AUDIO_PH * next(audio_iter), 1)
+       sample = sample.replace(AUDIO_PH, self.audio_token)
+   texts_inputs = self.tokenizer(tokenizer_text, padding=False, padding_side="left", ...)
+   batch = BatchFeature(data={**texts_inputs, **audio_inputs}, tensor_type=...)
+   ```
+5. **保留 Qwen2.5 spatial processor 的所有结构**：
+   - `_normalize_audio_list`：mono `[T] / [1,T] / [T,1]` 或 FOA `[4,T] / [T,4]` → `[C, T_audio]` 标准化
+   - `_truncate_audio_list`：>20s 截断 + warning
+   - `_build_spatial_audio_payload`：FOA padded 到 `[B, 320000, 4]`，attention mask `[B, 320000]`，spatial_token_lengths via `_samples_to_spatial_beats_tokens`（half-to-even rounding，2.5Hz target → 5s → 12 token、20s → 50 token）
+   - `_validate_text_modal_placeholders`：每条样本必须 `audio_count == 1`；FOA 必须 `spatial_count == 1`，mono 必须 `spatial_count == 0`；模态顺序 `<|AUDIO|> < <|spatial|>`
+   - `_expand_spatial_placeholders`：把单个 `<|spatial|>` 展开成 N 个，N = spatial_token_lengths[i]
+   - `sync_spatial_tokenizer_with_model`：调 `model.sync_spatial_tokenizer(tokenizer)`，把 token id 同步到 model.config
+
+**Smoke 测试**：
+
+```python
+proc = Qwen3OmniMoeSpatialProcessor(feature_extractor=fe, tokenizer=tok)
+foa = np.random.randn(4, 16000 * 5).astype(np.float32)  # 5s FOA
+text = '<|im_start|>user\n<|audio_start|><|AUDIO|><|audio_end|><|spatial|>Direction?<|im_end|>'
+batch = proc(text=text, audio=[foa], return_tensors='pt')
+# input_features=[1,128,500], spatial_audio=[1,320000,4], spatial_token_lengths=[12], spatial_audio_lengths=[80000]
+# input_ids 中 spatial token 数 = 12, audio_pad token 数 = 65（按 mel 帧数）
+```
+
+### 2.4 文件 3：`spatial_qwen/model/modeling_spatial_qwen3_thinker.py`（359 行）
+
+**目的**：Spatial-BEATs 注入版的 Qwen3 Thinker。
+
+**两个对外类**：
+
+| 类名 | 父类 | 用途 |
+|---|---|---|
+| `Qwen3OmniMoeSpatialThinkerForConditionalGeneration` | `Qwen3OmniMoeThinkerForConditionalGeneration`（fork） | 真正的 spatial 模型，所有逻辑在这里 |
+| `Qwen3OmniMoeSpatialForConditionalGeneration` | 上者 | **wrapper**：让 `model.thinker == model`，`model.disable_talker()` 是 no-op，方便复用 Qwen2.5 训练脚本 |
+
+**关键设计**：
+
+1. **`__init__`**：
+   - 父类 init 先建 `audio_tower / visual / model / lm_head` （Qwen3 的 Thinker），然后我们加 `spatial_beats_encoder`（`SpatialBEATsEncoderWrapper`）和 `spatial_beats_projector`（pixel-shuffle，output_dim=2048）。
+   - 验证 rate 一致性：`encoder_rate / shuffle_factor == target_rate`（默认 10/4 = 2.5）。
+   - 仅支持 `spatial_encoder_type == "spatial_beats"`，其他抛 `NotImplementedError`。
+
+2. **`forward`** 注入流程（与 Qwen2.5 spatial 一致，只改投影维度）：
+   ```
+   if 没有 spatial 输入: return super().forward(*args, **kwargs)
+   if use_audio_in_video: NotImplementedError
+   spatial_tokens = self._resolve_spatial_tokens(...)        # BEATs encoder forward
+   projected = self.spatial_beats_projector(spatial_tokens)   # pixel-shuffle，输出 [B, T_after, 2048]
+   if shuffle_factor > 1: 调整 spatial_token_lengths
+   projected, lengths = self._align_to_placeholders(...)      # 与 input_ids 中 <|spatial|> 数对齐
+   flat = self._flatten_projected_spatial(...)                # [sum_T, 2048]
+   inputs_embeds = self.get_input_embeddings()(input_ids)
+   spatial_mask = (input_ids == config.spatial_token_index).unsqueeze(-1).expand_as(inputs_embeds)
+   inputs_embeds = inputs_embeds.masked_scatter(spatial_mask, flat)
+   kwargs["inputs_embeds"] = inputs_embeds
+   return super().forward(*args, **kwargs)
+   ```
+
+3. **`get_rope_index` 不重写**！🎯 关键发现：
+   - Qwen3 上游 `get_rope_index`（`Qwen3OmniMoePreTrainedModelForConditionalGeneration` 类，line 260）对每个 token 检查是不是 `vision_start_token_id` / `audio_start_token_id` / `image_token_id` / `video_token_id`；任何**不**匹配特殊 id 的 token 都被当作普通 text，分配 sequential position id。
+   - 我们注册的 `<|spatial|>` 是新 id，不在父类已知 id 列表中 → 自动走 text 分支 → 拿到与文本一致的 sequential position id。
+   - mrope_section (24, 20, 20) + interleaved=True 由底层 `Qwen3OmniMoeThinkerTextRotaryEmbedding` 处理，对调用方透明。
+   - **结论：spatial token 的 RoPE 完全由父类正确处理，不需要任何重写**。这是相对 Qwen2.5 spatial 实现（重写了 200 行 get_rope_index）的最大简化。
+
+4. **`sync_spatial_tokenizer`**：把 `<|spatial|>` 加进 tokenizer，resize embeddings，把 token id 写回 `config.spatial_token_index`。与 Qwen2.5 实现完全一致。
+
+5. **`Qwen3OmniMoeSpatialForConditionalGeneration` wrapper**：
+   ```python
+   class Qwen3OmniMoeSpatialForConditionalGeneration(Qwen3OmniMoeSpatialThinkerForConditionalGeneration):
+       @property
+       def thinker(self):
+           return self
+       def disable_talker(self):
+           return None  # Qwen3 路径不构建 talker
+   ```
+   这样 Qwen2.5 训练脚本里所有 `model.thinker.spatial_beats_encoder.X` 访问、`model.disable_talker()` 调用都直接 work。
+
+6. **Talker 完全跳过**：30B 的 talker 占额外 ~3 GB，且训练脚本不需要它（QA 任务只需要 LM logits）。Wrapper 直接以 Thinker 作为 top-level model，与 fork 的顶层 `Qwen3OmniMoeForConditionalGeneration` 解耦（也避开了 fork 顶层 config 加载的 bug）。
+
+### 2.5 复用的现有模块（**完全不动**）
+
+| 模块 | 路径 | 作用 |
+|---|---|---|
+| `SpatialBEATsEncoderWrapper` | `spatial_qwen/modules/spatial_beats_encoder.py` | FOA → 768d encoder hidden @10Hz；checkpoint 加载逻辑 |
+| `SpatialTokenProjector` / `build_spatial_token_projector` | `spatial_qwen/modules/spatial_token_projector.py` | pixel_shuffle / mlp / mlp_ln 三种 projector，**通过 `output_dim=config.text_config.hidden_size` 自动适配 4096→2048** |
+
+### 2.6 不写的文件
+
+- ❌ `feature_extraction_qwen3_audio_spatial.py`：Qwen2.5 版本是 SPUR/banding 代码（用 SimplePowerVector 等），我们 BEATs 路径不需要它；Whisper 风格的 mel 由 `WhisperFeatureExtractor` 处理 mono branch，FOA 由 spatial processor 自己 pad，没有任何 banding。
+- ❌ `processing_qwen3_omni_moe.py` 的本地副本：直接 import fork 上游。
+
+---
+
+## 3. 阶段 2：训练入口 + shell（已完成）
+
+### 3.1 文件 4：`train_spatial_beats_qa_qwen3.py`（185 行）
+
+**策略**：**不复制** 1500 行的 `train_spatial_beats_qa.py`。改为 thin wrapper，只 monkey-patch 唯一两个 Qwen-version-specific 函数。
+
+```python
+import train_spatial_beats_qa as _trainer
+
+def _build_processor_qwen3(model_id, sqr):
+    tok = AutoTokenizer.from_pretrained(model_id)
+    fe  = AutoFeatureExtractor.from_pretrained(model_id)
+    return Qwen3OmniMoeSpatialProcessor(feature_extractor=fe, tokenizer=tok)
+
+def _build_model_qwen3(args, processor):
+    raw = json.load(open(os.path.join(args.model_id, "config.json")))
+    cfg = Qwen3OmniMoeSpatialThinkerConfig(**raw["thinker_config"])
+    cfg.spatial_encoder_type = "spatial_beats"
+    cfg.spatial_beats_checkpoint_path = args.beats_checkpoint
+    ...
+    cfg.text_config.router_aux_loss_coef = 0.0      # 关 MoE aux loss
+    cfg.text_config.output_router_logits = False
+    cfg.loss_type = "ForCausalLMLoss"
+    model = Qwen3OmniMoeSpatialForConditionalGeneration.from_pretrained(
+        args.model_id, config=cfg, torch_dtype=bf16, attn_implementation="sdpa", low_cpu_mem_usage=True,
+    )
+    processor.sync_spatial_tokenizer_with_model(model)
+    model.disable_talker()  # no-op
+    if args.gradient_checkpointing:
+        _trainer.enable_gradient_checkpointing(model)
+    enc = model.spatial_beats_encoder
+    if enc is not None:
+        enc._build_model()        # CPU 上构建，再迁到 projector 设备
+        if device_map is not None:
+            enc.to(next(model.spatial_beats_projector.parameters()).device)
+    return model
+
+def main():
+    _trainer.build_processor = _build_processor_qwen3
+    _trainer.build_model = _build_model_qwen3
+    return _trainer.main()
+```
+
+**为什么 monkey-patch 而不是 fork**：
+- `train_spatial_beats_qa.py` 的 1500 行里 99% 是 dataloader / DDP / LoRA / valid / checkpoint / resume，全部 backbone-agnostic。
+- 唯一两个 Qwen-specific 函数：`build_processor` 和 `build_model`，正好用 monkey-patch 替换。
+- 通过 wrapper class `Qwen3OmniMoeSpatialForConditionalGeneration` 提供 `model.thinker == model`，所有 `model.thinker.spatial_beats_encoder` 访问继续生效��
+
+**关键细节**：
+- 关 MoE router aux loss：`router_aux_loss_coef=0.0`、`output_router_logits=False`。我们只 LoRA attention 层，experts 都是冻结的，aux loss 没有任何意义；保持开启会污染 LM loss 信号。
+- `attn_impl="auto"` 自动 fallback 到 sdpa（spur-qwen3 没装 flash-attn）。
+- `low_cpu_mem_usage=True` 在 30B 模型上必开，避免双倍显存占用峰值。
+- BEATs 的 `_build_model()` 推迟到 model 加载之后再调，避开父类 init 顺序冲突。
+
+### 3.2 文件 5：`shell/launch_train_spatial_beats_qwen3.sh`（230 行）
+
+复制 `shell/launch_train_spatial_beats_v13d_easy.sh` 的 3 阶段结构（projector → encoder_lora → beats_lora），关键改动：
+
+```bash
+QWEN3_OMNI_FORK="${QWEN3_OMNI_FORK:-/apdcephfs_cq10/.../model/transformers/src}"
+export QWEN3_OMNI_FORK   # train script 的 sys.path bootstrap 依赖此环境变量
+
+MODEL_ID="${MODEL_ID:-/apdcephfs_cq12/share_302080740/model/Qwen3-Omni-30B-A3B-Instruct}"
+RUN_ROOT="${RUN_ROOT:-${ROOT_DIR}/runs/qwen3_v1_easy_llmqa}"   # 与 Qwen2.5 完全隔离
+
+BATCH_SIZE="${BATCH_SIZE:-1}"        # 30B vs 7B → 单 batch，省显存
+GRAD_ACCUM_STEPS="${GRAD_ACCUM_STEPS:-8}"   # 全局 batch = 1 * 8 * 8 = 64，与 v13d 对齐
+ATTN_IMPL="${ATTN_IMPL:-sdpa}"       # spur-qwen3 没装 flash-attn
+USE_GRADIENT_CHECKPOINTING="${USE_GRADIENT_CHECKPOINTING:-1}"   # 30B 推荐打开
+
+LORA_TARGET_MODULES=(${LORA_TARGET_MODULES:-q_proj k_proj v_proj o_proj})  # 只贴 attention，experts 不动
+
+# stage 学习率与 v13d 一致
+STAGE1_LR=5e-5  STAGE1_PROJECTOR_LR=1e-4
+STAGE2_LR=5e-5  STAGE2_LORA_LR=5e-5  STAGE2_PROJECTOR_LR=3e-5
+STAGE3_LR=3e-5  STAGE3_LORA_LR=3e-5  STAGE3_PROJECTOR_LR=1e-6  STAGE3_BEATS_LR=1e-6
+
+run_train() {
+  CUDA_VISIBLE_DEVICES="${GPUS}" QWEN3_OMNI_FORK="${QWEN3_OMNI_FORK}" \
+    torchrun --nnodes=... --nproc_per_node=${NPROC} ... \
+      "${ROOT_DIR}/train_spatial_beats_qa_qwen3.py" "$@"
+}
+```
+
+**前置检查**：BEATs ckpt / QA 数据 / Qwen3 模型目录 / fork 路径必须存在，缺一个就 exit 1。
+
+### 3.3 默认路径
+
+```
+QA_ROOT     = /apdcephfs_cq10/.../genQA/all_qa_llm_by_difficulty_v2/easy_filtered  (787K train / 84K valid / 84K test)
+BEATS_CKPT  = /apdcephfs_cq10/.../unilm/beats/checkpoints/spatial_beats_ov1_unified_v13d_exp/03_ov123_top4/best.pt
+BEATS_REPO  = /apdcephfs_cq10/.../unilm/beats
+RUN_ROOT    = runs/qwen3_v1_easy_llmqa
+                stage1_projector/
+                stage2_encoder_lora/
+                stage3_beats_lora/
+```
+
+### 3.4 调用范例
+
+```bash
+# 单机 8 卡，三阶段全流程
+bash shell/launch_train_spatial_beats_qwen3.sh
+
+# 从 stage2 开始
+START_STAGE=2 bash shell/launch_train_spatial_beats_qwen3.sh
+
+# smoke 测试（仅 32 样本）
+MAX_TRAIN_SAMPLES=32 STAGE1_EPOCHS=1 bash shell/launch_train_spatial_beats_qwen3.sh
+```
+
+---
+
+## 4. 阶段 3：DeepSpeed ZeRO-3（已完成 - asset，trainer 集成 TODO）
+
+### 4.1 显存预算
+
+| 组件 | 大小 |
+|---|---|
+| 30B BF16 权重 | 60 GB |
+| ZeRO-3 分片到 8 卡 | 7.5 GB / 卡 |
+| BEATs encoder（freeze） | 0.4 GB |
+| LoRA（attention only，r=16） | ~50M ≈ 0.1 GB trainable |
+| AdamW optimizer state（CPU offload） | 0 GB GPU |
+| Activations + KV cache（bs=1, seq~4096） | 10 GB / 卡 |
+| **预估每卡总占用** | **~22 GB**（40 GB 卡留 18 GB margin） |
+
+### 4.2 文件 6：`configs/ds_zero3_qwen3.json`（55 行）
+
+```json
+{
+  "bf16": {"enabled": true},
+  "zero_optimization": {
+    "stage": 3,
+    "offload_optimizer": {"device": "cpu", "pin_memory": true},
+    "offload_param":     {"device": "cpu", "pin_memory": true},
+    "overlap_comm": true,
+    "contiguous_gradients": true,
+    "stage3_gather_16bit_weights_on_model_save": true,
+    ...
+  },
+  "gradient_accumulation_steps": "auto",
+  "train_micro_batch_size_per_gpu": "auto",
+  "gradient_clipping": "auto",
+  "optimizer": {"type": "AdamW", "params": {"lr": "auto", ...}},
+  "scheduler": {"type": "WarmupDecayLR", "params": {...}}
+}
+```
+
+### 4.3 集成状态
+
+- ✅ JSON config 已写。
+- ❌ `train_spatial_beats_qa_qwen3.py` 还**没有** wire DeepSpeed。
+
+### 4.4 集成 TODO（约 1–2 天工作量）
+
+需要在 `train_spatial_beats_qa.py` 里加：
+
+```python
+import deepspeed
+
+# 替换 DDP wrap：
+# OLD: model = DistributedDataParallel(model, device_ids=[local_rank])
+# NEW:
+model_engine, optimizer, _, scheduler = deepspeed.initialize(
+    args=args, model=model, model_parameters=trainable_params,
+    config=os.environ.get("DS_CONFIG", "configs/ds_zero3_qwen3.json"),
+)
+
+# 替换训练步：
+# OLD: loss.backward(); optimizer.step(); optimizer.zero_grad()
+# NEW: model_engine.backward(loss); model_engine.step()
+
+# checkpoint：用 model_engine.save_checkpoint(...) 而不是 torch.save
+```
+
+### 4.5 临时回退（在 DS 集成完成前）
+
+- **stage 1 projector_only**：试 `BATCH_SIZE=1 GRAD_ACCUM_STEPS=8` + plain DDP；如果 30B forward 单卡 OOM，退到 `device_map="auto"` 单 replica（NPROC=1，HF accelerate 跨 8 卡 sharding，仅推理时验证过）。
+- **stage 2 / 3 LoRA + optimizer state**：必须 ZeRO-3，**集成完成前不要跑**。
+- 阶段 5 真实训练前，先把 DS 接入做完。
+
+---
+
+## 5. 阶段 4：bench（已完成）
+
+### 5.1 文件 7：`scripts/bench_test_generate_qwen3.py`（66 行）
+
+Thin wrapper，monkey-patch 三层：
+
+```python
+import train_spatial_beats_qa as _trainer
+from train_spatial_beats_qa_qwen3 import _build_processor_qwen3, _build_model_qwen3
+
+# 1. patch trainer module
+_trainer.build_processor = _build_processor_qwen3
+_trainer.build_model     = _build_model_qwen3
+
+# 2. patch batch_bench module（已经 from train_spatial_beats_qa import build_processor）
+from scripts import batch_bench_spatial_beats_qa as _bb
+_bb.build_processor = _build_processor_qwen3
+_bb.build_model     = _build_model_qwen3
+
+# 3. patch bench_test_generate（已经 from batch_bench import instantiate_model_for_checkpoint）
+from scripts import bench_test_generate as _bench
+_bench.build_processor = _build_processor_qwen3
+_bench.build_model     = _build_model_qwen3
+
+def main(): return _bench.main()
+```
+
+**为什么必须三层 patch**：Python `from M import name` 是绑定**当前时刻**的引用，不是 lazy lookup。`bench_test_generate.py` 在 import 时已经把 `build_processor` 复制到自己的 module namespace；只 patch `train_spatial_beats_qa.build_processor` 不够，必须把所有「捕获了引用的 module」都重新绑定。
+
+### 5.2 文件 8：`scripts/run_bench.py`（修改 1 处）
+
+加 2 个 baseline 路由：
+
+```python
+BASELINE_TO_MODULE = {
+    "spatial-qwen":       ("scripts.bench_test_generate",       []),
+    "spatial-qwen3":      ("scripts.bench_test_generate_qwen3", []),     # ← 新
+    "zero-spatial":       ("scripts.bench_test_generate",       ["--spatial-ablation", "zero"]),
+    "zero-spatial-qwen3": ("scripts.bench_test_generate_qwen3", ["--spatial-ablation", "zero"]),  # ← 新
+    "iv":                 ("scripts.bench_test_generate_iv",    []),
+    "neural-iv":          ("scripts.bench_test_generate_iv",    []),
+    "spatial-flamingo":   ("scripts.bench_test_generate_af3",   []),
+}
+```
+
+### 5.3 调用范例
+
+```bash
+# Qwen3 baseline 实测
+torchrun --nproc_per_node=8 scripts/run_bench.py \
+    --baseline spatial-qwen3 \
+    --checkpoint-paths runs/qwen3_v1_easy_llmqa/stage2_encoder_lora/checkpoints/best_trainable.pt \
+    --qa-root /apdcephfs_cq10/.../easy_filtered \
+    --split test --batch-size 1 --num-workers 4 \
+    --output-dir runs/qwen3_v1_easy_llmqa/stage2_encoder_lora/bench/test
+
+# 关 spatial 通道的 ablation
+torchrun --nproc_per_node=8 scripts/run_bench.py \
+    --baseline zero-spatial-qwen3 \
+    --checkpoint-paths .../best_trainable.pt \
+    --qa-root .../easy_filtered --split test
+```
+
+bench 完成后用既有 `scripts/score_test_predictions.py` 打分（CPU only，可加 LLM judge）。
+
+---
+
+## 6. 完整文件清单
+
+### 6.1 新建的文件（10 个，~1300 行）
+
+| # | 路径 | 行数 | 状态 | 说明 |
+|---|---|---|---|---|
+| 1 | `spatial_qwen/model/configuration_qwen3_omni.py` | 202 | ✅ smoke 通过 | spatial Thinker config |
+| 2 | `spatial_qwen/model/processing_spatial_qwen3.py` | 494 | ✅ smoke 通过 | FOA + `<\|spatial\|>` processor |
+| 3 | `spatial_qwen/model/modeling_spatial_qwen3_thinker.py` | 359 | ✅ 编译通过 | spatial 注入 Thinker + wrapper |
+| 4 | `train_spatial_beats_qa_qwen3.py` | 185 | ✅ `--help` 通过 | 训练 entrypoint（monkey-patch 包装） |
+| 5 | `shell/launch_train_spatial_beats_qwen3.sh` | 230 | ✅ 前置检查通过 | 三阶段 launcher |
+| 6 | `configs/ds_zero3_qwen3.json` | 55 | ✅ JSON 合法 | DS ZeRO-3 配置 asset |
+| 7 | `scripts/bench_test_generate_qwen3.py` | 66 | ✅ `--help` 通过 | bench 入口（monkey-patch） |
+| 8 | `docs/qwen3_env.md` | ~80 | ✅ | 环境笔记 + memory budget |
+| 9 | `docs/spatial_qwen3_design.md` | 本文 | ✅ | **本设计文档** |
+| 10 | `tests/test_spatial_qwen3_pipeline.py` | — | ⏳ TODO（阶段 1d） | pytest 集成测试 |
+
+### 6.2 修改的文件（1 个）
+
+| 路径 | 改动 |
+|---|---|
+| `scripts/run_bench.py` | `BASELINE_TO_MODULE` 加 2 项：`spatial-qwen3` / `zero-spatial-qwen3` |
+
+### 6.3 完全不动的文件
+
+```
+所有 Qwen2.5 路径文件（≈ 12 个），包括：
+    train_spatial_beats_qa.py
+    spatial_qwen/model/modeling_spatial_thinker.py
+    spatial_qwen/model/processing_spatial.py
+    spatial_qwen/model/configuration.py
+    spatial_qwen/model/modeling_qwen2_5_omni.py / modular_qwen2_5_omni.py / processing_qwen2_5_omni.py
+    spatial_qwen/model/feature_extraction_qwen2_audio_spatial.py
+    shell/launch_train_spatial_beats_v13d_easy.sh
+
+所有 IV / Neural-IV 路径文件（≈ 6 个）
+
+所有 AF3 路径文件（≈ 4 个）
+
+所有 backbone-agnostic 模块（complete share between Qwen2.5 / Qwen3）：
+    spatial_qwen/modules/spatial_beats_encoder.py
+    spatial_qwen/modules/spatial_token_projector.py
+    spatial_qwen/modules/iv_spatial_adapters.py
+    spatial_qwen/modules/seldnet233_*.py
+    spatial_qwen/encoders/seldnet/  / encoders/beats/
+    scripts/score_test_predictions.py
+    scripts/precompute_*.py
+
+spur conda env 本身（保持 Qwen2.5 训练环境完全不变）
+所有 runs/v13d_easy_llmqa/ 等历史 ckpt
+```
+
+---
+
+## 7. 运行策略 / 复现指南
+
+### 7.1 完整运行清单（按时间顺序）
+
+#### Step 0：环境准备（已完成，仅记录）
+
+```bash
+# 0a. 克隆 conda env
+conda create -n spur-qwen3 --clone spur
+
+# 0b. fork editable install
+conda activate spur-qwen3
+cd /apdcephfs_cq10/share_1603164/user/schmittzhu/model/transformers
+pip install -e .
+
+# 0c. 修兼容性
+pip install -U "huggingface-hub>=1.3.0,<2.0"     # → 1.14.0
+pip install -U "tokenizers>=0.22.0,<=0.23.0"     # → 0.22.2
+
+# 0d. 拍快照
+cp -r /apdcephfs_cq10/.../model/transformers /apdcephfs_cq10/.../model/transformers.snapshot-pre-qwen3
+
+# 0e. 验证
+python -c "
+import sys; sys.path.insert(0, '/apdcephfs_cq10/.../model/transformers/src')
+from transformers.models.qwen3_omni_moe import (
+    Qwen3OmniMoeThinkerForConditionalGeneration,
+    Qwen3OmniMoeProcessor,
+    Qwen3OmniMoeThinkerConfig,
+)
+print('OK')
+"
+```
+
+#### Step 1：单元烟测（阶段 1d，TODO）
+
+```bash
+conda activate spur-qwen3
+cd /apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen
+pytest tests/test_spatial_qwen3_pipeline.py -v
+```
+
+测试内容（待写）：
+1. 配置 from_pretrained roundtrip 保住 spatial 字段
+2. processor 对单条 FOA 5s 输入 → spatial_token_lengths == 12，audio token 数与 mel 长度一致
+3. 用 tiny config（hidden=256, layers=2, experts=4）实例化 model，random init forward 一次，logits shape OK，无 NaN
+4. `model.thinker is model`，`disable_talker()` 不报错
+
+#### Step 2：smoke 训练（阶段 5，第一步）
+
+```bash
+# 32 样本 stage1 projector_only，确认 loss 5 步内单调下降
+conda activate spur-qwen3
+cd /apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen
+MAX_TRAIN_SAMPLES=32 STAGE1_EPOCHS=1 BATCH_SIZE=1 GRAD_ACCUM_STEPS=2 \
+NPROC=1 \
+bash shell/launch_train_spatial_beats_qwen3.sh
+```
+
+> 单卡 30B 极可能 OOM，需要先把 DS 接入或先用 `device_map=auto`。**这是阶段 3+5 的真正接入点**。
+
+#### Step 3：DeepSpeed 接入（阶段 3 finalize）
+
+按 §4.4 在 `train_spatial_beats_qa.py` 加 `deepspeed.initialize` 调用，`shell/launch_train_spatial_beats_qwen3.sh` 加 `--deepspeed configs/ds_zero3_qwen3.json` arg 透传，runtime 用 `deepspeed --num_gpus=8 train_spatial_beats_qa_qwen3.py ...`（或 torchrun + DS 自动检测）。
+
+#### Step 4：真实 stage1（阶段 5 主体）
+
+```bash
+# 8 卡 ZeRO-3，1 epoch easy_filtered (~787K samples → ~12K opt step)
+conda activate spur-qwen3
+START_STAGE=1 STAGE1_EPOCHS=1 \
+BATCH_SIZE=1 GRAD_ACCUM_STEPS=8 \
+bash shell/launch_train_spatial_beats_qwen3.sh
+```
+
+预期墙钟时间：~24 h / epoch（参考 Qwen2.5-7B v13d 是 ~6 h，30B + ZeRO-3 offload ~4× slower）。
+
+#### Step 5：bench（阶段 4 应用）
+
+```bash
+torchrun --nproc_per_node=8 scripts/run_bench.py \
+    --baseline spatial-qwen3 \
+    --checkpoint-paths runs/qwen3_v1_easy_llmqa/stage1_projector/checkpoints/best_trainable.pt \
+    --qa-root /apdcephfs_cq10/.../easy_filtered \
+    --split test --batch-size 1 --num-workers 4
+
+# 拿到 predictions.jsonl 后打分（CPU）
+python scripts/score_test_predictions.py \
+    --predictions-jsonl runs/qwen3_v1_easy_llmqa/stage1_projector/bench/test/best/predictions.jsonl \
+    --qa-root /apdcephfs_cq10/.../easy_filtered --split test \
+    --llm-judge --llm-concurrency 8
+```
+
+#### Step 6：隔离验证（不该破坏老 Qwen2.5 路径）
+
+```bash
+# 切回 spur env，跑老 v13d 推理，对照已有 score_result.json
+conda activate spur
+bash shell/bench_easy_all.sh BASELINES=spatial-qwen
+# 期望：score_result.json 的指标与之前完全一致
+```
+
+### 7.2 调试 / 故障排查
+
+| 现象 | 排查 |
+|---|---|
+| import 时 `qwen3_omni_moe` 找不到 | `echo $QWEN3_OMNI_FORK`；fork editable install 是否在当前 env：`pip show transformers \| grep Location` |
+| `Qwen3OmniMoeConfig.from_pretrained` 崩溃 | **不要**用顶层 config；按 §1.5.1 走 raw JSON |
+| processor `__init__` 报 image_processor 类型错 | 不要直接 `Qwen3OmniMoeProcessor(...)`，用我们 subclass `Qwen3OmniMoeSpatialProcessor` |
+| `<\|spatial\|>` 不在 vocab | processor 构造时会自动 `add_special_tokens`；如果用了不同 tokenizer，重跑 `proc.sync_spatial_tokenizer_with_model(model)` |
+| forward 报 spatial token 数对不上 | 见 `_validate_spatial_mask_count`：placeholder count vs projected token count；通常是 shuffle_factor 计算不一致或 prompt 漏了 `<\|spatial\|>` |
+| MoE router 出现 NaN aux loss | 我们已在 build_model 里 `router_aux_loss_coef=0.0`；如果 fork 行为变了重新 audit |
+| 30B 单卡 OOM | 必须 ZeRO-3 + offload；先确认 DS wire 已完成（§4.4） |
+| spur 老训练受影响 | 不应该，所有 Qwen2.5 文件都没动；如果 spur env 里 `import transformers` 也变成 5.0.0.dev0，说明 pip 装错 env 了 |
+
+### 7.3 路径锚点
+
+```
+仓库根                : /apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen
+transformers fork    : /apdcephfs_cq10/share_1603164/user/schmittzhu/model/transformers/src
+fork 快照             : /apdcephfs_cq10/share_1603164/user/schmittzhu/model/transformers.snapshot-pre-qwen3
+Qwen3-30B 权重       : /apdcephfs_cq12/share_302080740/model/Qwen3-Omni-30B-A3B-Instruct
+BEATs ckpt (v13d)   : /apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/checkpoints/spatial_beats_ov1_unified_v13d_exp/03_ov123_top4/best.pt
+QA easy_filtered    : /apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2/easy_filtered
+RUN_ROOT (Qwen3)    : runs/qwen3_v1_easy_llmqa
+RUN_ROOT (Qwen2.5)  : runs/v13d_easy_llmqa  ← 不动
+```
+
+### 7.4 工程不变量（绝不可破坏）
+
+1. spur conda env 不能装新包；新包只装 spur-qwen3。
+2. 任何 Qwen2.5 / IV / AF3 文件不修改（包括 shell）。
+3. `runs/v13d_easy_llmqa/` 等历史 ckpt 不被覆盖。
+4. `scripts/run_bench.py` 的修改只是新增 dict 项，旧 baseline 行为完全保持。
+5. fork 自身不修改；如果发现 fork bug 必须改，先在 spur-qwen3 私有 fork copy 上改，记录在本文档 §1.5。
+
+---
+
+## 8. 风险登记 / 回滚
+
+| 风险 | 触发概率 | 缓解 |
+|---|---|---|
+| fork 上游 API 与我们的 subclass 假设不一致 | 中 | 阶段 1d pytest 覆盖关键 path（forward shape、generate、masked_scatter） |
+| MoE 训练梯度不稳 | 低-中 | 已关 aux loss；stage1 lr 沿用 v13d 的 5e-5 保守值；首次 32-sample smoke 必跑 |
+| ZeRO-3 + 自定义 spatial 模块（不在 DS 视野里）的状态管理 | 中 | spatial encoder 走 `model.spatial_beats_encoder` 直接挂在 thinker 上，DS Z3 会自动 partition；如果有问题，按 train_spatial_beats_qa.py 已有的 unfreeze_modules 模式手动 register |
+| 30B + 8 卡 + offload 一个 epoch 跑不完 | 中-高 | 阶段 5 先用 `--max-train-samples=50000` 测节奏，看 24h 跑不跑得完 |
+| pip 升级污染了 spur env | 低（已隔离） | 一律先 `conda activate spur-qwen3`；若发生用 `pip install -r requirements.txt --force-reinstall` 修 spur |
+| fork 快照与本地 fork 在我不知情的情况下分叉 | 低 | 阶段 5 跑前 `diff -r transformers/src transformers.snapshot-pre-qwen3/src \| head` 检查 |
+
+### 回滚步骤
+
+1. 删除 spur-qwen3：`conda env remove -n spur-qwen3`
+2. 删除新建的 10 个文件（见 §6.1）
+3. revert `scripts/run_bench.py` 的 1 处改动（删除 `spatial-qwen3` / `zero-spatial-qwen3` 两个 dict 项）
+4. 用快照恢复 fork（仅当确实改坏过 fork）：`rm -rf transformers/src && cp -r transformers.snapshot-pre-qwen3/src transformers/`
+
+---
+
+## 9. 验收清单
+
+| 阶段 | 验收点 | 状态 |
+|---|---|---|
+| 0 | conda env list 里有 spur-qwen3，`from transformers.models.qwen3_omni_moe import ...` 成功 | ✅ |
+| 1a | `Qwen3OmniMoeSpatialThinkerConfig` 从真实 30B config.json 实例化 + roundtrip 保留所有 spatial 字段 | ✅ |
+| 1b | spatial processor 对 FOA 5s 输入产出 12 spatial token + 65 audio token，mixed-batch padding 正确 | ✅ |
+| 1c | spatial Thinker subclass 静态 import 通过，wrapper class `model.thinker == model` | ✅（forward 实测留 1d） |
+| 1d | pytest tests/test_spatial_qwen3_pipeline.py 全部通过 | ⏳ TODO |
+| 2 | `python train_spatial_beats_qa_qwen3.py --help` 输出完整 arg 列表 | ✅ |
+| 2 | shell 前置检查（路径存在）通过 | ✅ |
+| 3 | configs/ds_zero3_qwen3.json 合法 | ✅ |
+| 3 | DS 接入 trainer（engine.backward / engine.step） | ⏳ TODO |
+| 4 | `python scripts/run_bench.py --help` 显示 `spatial-qwen3` / `zero-spatial-qwen3` 选项 | ✅ |
+| 4 | `python scripts/bench_test_generate_qwen3.py --help` 输出完整 arg | ✅ |
+| 5 | smoke：32 样本 stage1 projector_only 5 步 loss 单调下降 | ⏳ TODO |
+| 5 | 真实：1 epoch easy_filtered stage1 训练完成，valid loss 收敛 | ⏳ TODO |
+| 隔离 | `bash shell/bench_easy_all.sh BASELINES=spatial-qwen` 在 spur env 下结果与之前一致 | ⏳ 跑过即可 |
+
+---
+
+## 10. 备忘 / 延展
+
+- **IV / Neural-IV / SELD233 在 Qwen3 上的支持**：本期不做。等 Spatial-BEATs 验证通过后再给 modeling_spatial_qwen3_thinker.py 加分支（结构与 Qwen2.5 spatial 一致，区别只是 projector output_dim=2048）。
+- **Talker 接入**：长期来看 talker 能让 Qwen3 输出语音回答；非空间 QA 任务用不上，暂不做。
+- **flash-attn**：本机 CUDA 12.8 装 flash-attn 2.x 编译耗时 ~30 min，对 30B prompt+spatial ~1k token 的注意力大概给 +15% 吞吐。值得装但不阻塞 stage5。
+- **`device_map="auto"` + ZeRO-3 互斥**：`from_pretrained` 时不要同时传两者；DS 内部会处理 sharding。
+- **更长上下文 / 更多 spatial token**：当前 50 token / 20 s clip。若要扩�� 1 min 音频（150 token），需要改 `spatial_beats_max_audio_seconds=60.0`，processor side `_truncate_audio_list` 会自动 truncate；attention 上 50 vs 150 token 对 30B 来说仍可忽略。
+
+---
+
+**文档结束**

docs/spur_seld_handover.md

@@ -0,0 +1,312 @@
+# SPUR SELD / Encoder-Only Handover Summary
+
+## 1. Goal Evolution
+The work went through two phases:
+
+1. `encoder-only probing / pretraining`
+- Train only SPUR spatial frontend + encoder
+- Use lightweight heads
+- Verify whether encoder tokens contain usable spatial information
+- This line is useful as a probe / pretraining result
+
+2. `paper-quality main result`
+- The current conclusion is that the custom encoder-only line is not strong enough to be the paper's main SELD result
+- The recommended main-result route is:
+  - keep `DCASE2024_seld_baseline` official head / ADPIT / decoding / metrics
+  - replace only the backbone with SPUR
+
+## 2. Repos Involved
+Main repo:
+- `/apdcephfs_cq10/share_1603164/user/schmittzhu/code/spur-qwen-2.5-omni`
+
+Reference repos:
+- `/apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline`
+- `/apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-AST`
+
+## 3. Data Understanding
+### STARSS23 metadata
+File:
+- `/apdcephfs_cq10/share_1603164/user/schmittzhu/data/metadata/STARSS23.json`
+
+Current metadata semantics:
+- `trajectory.npy` is `active-only`, not full-length with NaNs
+- static sources:
+  - `DoA` and `distance` are directly present
+  - `trajectory` is nearly constant
+- dynamic sources:
+  - `DoA=[]`
+  - `distance=""`
+  - but `trajectory.npy` contains per-frame `az/el/dist`
+
+Important conclusion:
+- data itself is not broken
+- but random chunk sampling originally did **not** guarantee event-rich chunks
+- this was later changed to `event-aware chunk sampling`
+
+### STARSS23 concurrency stats (already computed)
+Per-scene total concurrency roughly:
+- `P90 = 3`
+- `P95 = 3`
+- `P99 = 4`
+
+Per-class concurrency roughly:
+- `P90 = 2`
+- `P95 = 3`
+- `P99 = 3`
+- `max = 4`
+
+This supported `per-class K=3` for the encoder-only ADPIT line.
+
+### Scene durations
+STARSS23 scenes are long.
+Previously measured:
+- shortest about `32.5s`
+- mean about `157.9s`
+- median about `141.0s`
+- longest about `569.2s`
+
+This is why whole-scene training OOMed and chunk training was introduced.
+
+## 4. Encoder-Only Line That Was Implemented
+A custom encoder-only training line was built inside this repo.
+
+### Core architecture
+- `SpurSpatialFeatureExtractor`
+- `SpurSpatialEncoder`
+- lightweight heads for:
+  - SED/activity
+  - multi-ACCDOA
+  - distance (`logr`)
+
+### Time axis
+Final version for this line used:
+- `10 Hz`
+- strict axis alignment to labels
+- no interpolation in the training target definition
+
+### Key files added / modified
+Main training line:
+- `training-qwen-omni/spatial_encoder_adpit_dataset.py`
+- `training-qwen-omni/spatial_encoder_adpit_collator.py`
+- `training-qwen-omni/adpit_loss.py`
+- `training-qwen-omni/spatial_encoder_adpit_metrics.py`
+- `training-qwen-omni/train_spatial_encoder_adpit.py`
+- `qwen2_5_omni_spur/modules/spatial_encoder_adpit_model.py`
+- `examples/pretrain/spatial_encoder_adpit_10hz.yaml`
+- `scripts/check_spatial_encoder_adpit_batch.py`
+- `scripts/eval_spatial_encoder_probe.py`
+- `scripts/run_spatial_encoder_sweep.py`
+- `examples/pretrain/spatial_encoder_adpit_sweep.yaml`
+
+### Important implementation details
+- active-only trajectories are mapped back to full 10 Hz timeline
+- event-aware chunk sampling was added later
+- class-wise ADPIT target formatting was changed to be closer to DCASE style
+- official-style DCASE metrics were integrated
+- train/val split was changed from fold-based to scene-random split (`143/25` style) for convenience in this probe line
+- later debug JSON dumping was added for eval samples
+
+## 5. UFB / Frontend Situation
+Important context:
+- local dev environment often could not import the intended UFB path directly
+- a mel fallback path was temporarily used for local debugging / validation
+- on the real training machine, importing from merged `transformers` worked:
+  - `from transformers.models.qwen2_5_omni.modules.spur_spatial_features import SpurSpatialFeatureExtractor`
+  - `_use_ufb == True`
+
+We changed the custom training builder so it tries to import from merged `transformers` first when available.
+
+Still, for final paper experiments, relying on the official DCASE pipeline plus SPUR backbone is preferred over expanding this custom line further.
+
+## 6. Main Findings From Encoder-Only Experiments
+### What was demonstrated
+The encoder-only line did show that SPUR encoder outputs contain usable spatial information:
+- matched predictions had non-random DOA
+- distance predictions were not meaningless
+- the system was no longer in an all-zero / degenerate regime after several fixes
+
+### But the line is not good enough for main result
+Main persistent problem:
+- SED / class-activity prediction was much weaker than desired
+- F-score stayed very low
+- often severe class bias / false positives / false negatives
+
+In other words:
+- this line is acceptable as `encoder probe / pretraining evidence`
+- it is not a good main SELD system for the paper
+
+## 7. Important Bugs / Fixes That Happened
+### 7.1 Time-axis mismatch bug
+Problem:
+- GT token count computed by `round(duration * token_rate)`
+- model produced STFT-frame-count-based length
+- mismatch by 1 frame on some samples
+
+Fix:
+- dataset token count switched to the exact STFT/frame formula
+- dataset got `hop_length`, `win_length`, `stft_center` from config
+
+### 7.2 OOM on full scenes
+Problem:
+- long scenes + covariance-like frontend tensors caused OOM
+
+Fix:
+- introduced random `10-15s` chunk training
+- later changed to event-aware chunk sampling
+
+### 7.3 Dist loss was zero / not learning
+Problem:
+- early custom ADPIT implementation not fully aligned with official DCASE logic
+- winning pattern selection and distance supervision were not stable
+
+Fix:
+- moved toward fixed DCASE-style ADPIT patterns
+- later integrated explicit `loss_sed + loss_accdoa + loss_dist`
+
+### 7.4 Track-level SED experiment
+We tried changing class-level SED to track-level SED aligned with ADPIT tracks.
+
+Result:
+- this did **not** improve the custom line
+- likely made classification / activity learning less stable
+- final judgment: not a good direction for the custom probe line's main result
+
+## 8. Debug Instrumentation Added
+The user explicitly asked to add debug outputs to understand why predicted classes differ so much from GT.
+
+Implemented in:
+- `training-qwen-omni/spatial_encoder_adpit_metrics.py`
+- `training-qwen-omni/train_spatial_encoder_adpit.py`
+- config flags in `examples/pretrain/spatial_encoder_adpit_10hz.yaml`
+
+### Debug output path
+- `outputs/spatial_encoder_adpit_10hz/debug/epoch_XXX_eval_debug.json`
+
+### Debug contents
+For first N eval batches, the JSON contains:
+- GT vs Pred class lists
+- false positive classes
+- missed classes
+- top SED classes
+- top score classes
+- per-sample metadata:
+  - `sample_id`
+  - `scene_id`
+  - `audio_path`
+  - `start_token`
+  - `end_token`
+  - `valid_tokens`
+- per-class stats:
+  - `gt_frame_count`
+  - `pred_frame_count`
+  - `matched_frame_count`
+  - `fp_frame_count`
+  - `fn_frame_count`
+  - `gt_instance_count`
+  - `pred_instance_count`
+
+### Debug analysis already done
+The latest debug JSON showed a strong class bias:
+
+Frequent false-positive classes:
+- `Domestic_sounds`
+- `Music`
+- `Female_speech_and_woman_speaking`
+- `Male_speech_and_man_speaking`
+- `Musical_instrument`
+
+Frequent false-negative / poorly learned classes:
+- `Water_tap_and_faucet`
+- `Walk_and_footsteps`
+- `Bell`
+- `Knock`
+- `Laughter`
+
+Interpretation:
+- the custom line learned a biased class prior
+- low `F` is mainly caused by class-level SED bias and miscalibration
+- DOA is not the main bottleneck
+
+## 9. Current Assessment of Custom Encoder-Only Line
+### Useful for
+- encoder probing
+- pretraining evidence
+- showing SPUR tokens contain spatial information
+
+### Not suitable for
+- final paper main result
+- fair direct comparison against DCASE baseline as a complete SELD system
+
+### Best achieved regime in custom line
+Depending on run, best custom-line checkpoints reached roughly:
+- `F` only a few percent (`~0.03 to 0.07` range)
+- `AngE` around `30-40°`
+- `DistE` around `0.29-0.45`
+- still much worse than desired baseline-level detection performance
+
+The user explicitly noted baseline-level references such as:
+- baseline around `13%` F
+- best results around `54%` F
+
+We agreed the custom line should not be forced into the main-result role anymore.
+
+## 10. Agreed Strategic Direction
+The user agreed that the next main-result route should be:
+
+> integrate SPUR backbone into `DCASE2024_seld_baseline`
+
+This means:
+- keep official DCASE2024 data flow, target construction, ADPIT loss, decoding, metrics
+- replace only the backbone with SPUR
+
+Why this is preferred:
+1. fairer comparison
+2. detection / decoding path is already validated
+3. isolates SPUR backbone contribution
+4. stronger paper argument
+
+## 11. Recommended Next Task in New Conversation
+The next conversation should focus on:
+
+### Main-result integration plan
+Target repo:
+- `/apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline`
+
+Suggested approach:
+1. inspect official model input/output requirements
+2. identify where the baseline backbone lives
+3. create a new model variant that uses:
+   - `SpurSpatialFeatureExtractor`
+   - `SpurSpatialEncoder`
+4. add a projection/adaptation layer if needed
+5. keep official:
+   - label format
+   - `MSELoss_ADPIT`
+   - decoding
+   - metrics
+6. run baseline-compatible training with SPUR backbone
+
+### Avoid in next conversation
+- do not continue pushing the current custom encoder-only line as the main result line
+- do not keep iterating on custom SED head designs unless strictly for probe/ablation purposes
+
+## 12. If Continuing the Custom Line Anyway
+If the user still wants to use the current custom line for analysis only, the most relevant future improvement mentioned was:
+- add `per-class pos_weight` for SED
+- maybe per-class threshold sweep
+
+But this was explicitly judged lower priority than moving to the official DCASE2024 baseline route for the paper's main result.
+
+## 13. User Preferences / Constraints
+- Prefer concise, factual communication in Chinese
+- User is comfortable running experiments on another machine with merged `transformers`
+- User wants debugging outputs when investigating poor predictions
+- User cares about paper-quality experiments, not just toy success
+
+## 14. Immediate Recommendation to Future Assistant
+Do this next:
+1. analyze `DCASE2024_seld_baseline` model / data / loss interfaces
+2. produce a minimal SPUR-backbone integration plan
+3. then implement the baseline-compatible SPUR model
+
+Do **not** spend more time polishing the current custom encoder-only line as the main result system.

docs/unified_foa_fsd63_dataset.md

@@ -0,0 +1,255 @@
+# unified_spatial_foa_fsd63_all 数据集说明
+
+## 概述
+
+`unified_spatial_foa_fsd63_all` 是一批大规模统一格式的空间 FOA 数据集，将多个子集整合为统一的 schema（`spatial_foa_scene_v1`）。数据位于：
+
+```
+/apdcephfs_cq12/share_302080740/user/schmittzhu/data/unified_spatial_foa_fsd63_all/
+```
+
+**规模（scene/clip 数）：**
+
+| split | 总计     | sim_static | qa_sim  | dcase_real |
+|-------|----------|------------|---------|------------|
+| train | 329,610  | 250,761    | 63,439  | 15,410     |
+| valid | 35,093   | 25,158     | 5,303   | 4,632      |
+| test  | 35,237   | 28,507     | 6,149   | 581        |
+| **合计** | **399,940** | **304,426** | **74,891** | **20,623** |
+
+source 实例总数：**865,965**（含多源场景）。
+
+---
+
+## 目录结构
+
+```
+unified_spatial_foa_fsd63_all/
+├── train.jsonl                  # 329,610 行，每行一个 scene record
+├── valid.jsonl                  # 35,093 行
+├── test.jsonl                   # 35,237 行
+├── label_mapping.json           # 63 类 FSD 类别映射
+├── summary.json                 # 整体统计
+└── annotations/                 # 逐 scene CSV 标注
+    ├── train/
+    │   ├── sim_static/          # 仿真静态场景
+    │   ├── qa_sim/              # 仿真问答场景（lr_pair）
+    │   └── dcase_real/          # 真实录音子集
+    │       ├── realfoa/         # 自录 FOA（含 inf elevation / -1 distance）
+    │       ├── starss22/
+    │       ├── starss23/
+    │       ├── tau2019/
+    │       ├── tau2020/
+    │       ├── tau2021/
+    │       ├── tau2636586/
+    │       └── tau2636594/
+    ├── valid/  （同结构）
+    └── test/   （同结构）
+```
+
+---
+
+## JSONL Schema（每行字段）
+
+```jsonc
+{
+  "schema_version": "spatial_foa_scene_v1",
+  "split": "train",                        // "train" / "valid" / "test"
+  "data_source": "sim_static",             // 见下文子集说明
+  "dataset": "hm3d",                       // 细分数据集名（如 hm3d, starss22, realfoa …）
+  "scene_id": "train/hm3d/…/000016-foa__354626",
+
+  "audio": {
+    "foa_path": "/path/to/foa.wav",        // 4ch B-format FOA 音频，16 kHz
+    "duration_seconds": 7.37675
+  },
+
+  "scene_annotation_csv_path": "/path/to/scene.csv",  // 场景级 10Hz 逐帧 CSV（见下）
+
+  "scene_meta": {
+    "listener_position_cm": [x, y, z],    // 仿真子集有值，真实子集为 null
+    "receiver_yaw_deg": 140.99,
+    "room": "00009-vLpv2VX547B"
+  },
+
+  "sources": [                             // 每个音源的元信息列表
+    {
+      "label": "wind_instrument",          // FSD63 类别名
+      "label_id": 59,                      // 类别 ID（0–62）
+      "track_id": 0,                       // 对应 CSV 中的 track_id
+      "source_id": "src00",
+      "active_duration_seconds": 4.05,
+      "active_times": [[1.0, 5.05]],       // 活跃时间段列表
+      "direction_start": ["front","left","up"],  // 粗粒度方向（可为 null）
+      "direction_end":   ["front","left","up"],
+      "motion": {
+        "is_moving": false,
+        "pattern": null,
+        "description": null,
+        "scene_description": null
+      },
+      "mono_audio_path": "/path/to/mono.wav",    // 仿真子集有，真实子集为 null
+      "per_source_foa_path": "…",
+      "rir_path": "…",
+      "source_trajectory_csv_path": "…"          // 单源逐帧 CSV（可选用）
+    }
+  ],
+  "task_type": null
+}
+```
+
+### 与旧 JSONL 格式（`load_segments()` 期望格式）的对比
+
+| 旧格式字段   | 新格式对应位置                     |
+|------------|----------------------------------|
+| `foa_path` | `audio.foa_path`                 |
+| `csv_path` | `scene_annotation_csv_path`      |
+| `length`   | `audio.duration_seconds`         |
+| `stem`     | `scene_id`（需截取最后段作 stem）  |
+| `split`    | `split`（同名）                   |
+| `dataset`  | `data_source`（或 `dataset`）     |
+
+> **注意**：新格式与 `generate_segment_qa_pairs.base.load_segments()` **不兼容**，后者期望 `foa_path`、`csv_path`、`stem`、`length` 为扁平键。接入 QA 生成脚本时需要新增适配函数（参见 [待办事项](#待办事项)）。
+
+---
+
+## 场景级 CSV 格式
+
+与 DCASE 2024 SELD 基线格式**相同**（无 header 行，6 列，10 Hz 逐帧）：
+
+```
+frame_idx,class_id,track_id,azimuth_deg,elevation_deg,distance_cm
+```
+
+**示例（sim_static，完整标注）：**
+```
+10,59,0,13.934,11.313,184
+11,59,0,13.934,11.313,184
+```
+
+**示例（dcase_real/realfoa，部分缺失标注）：**
+```
+49,17,5,111.000,-inf,-1
+75,17,5,111.000,inf,-1
+```
+
+> **重要**：CSV 文件**无 header 行**，可直接被 `parse_csv_rows()` 按现有逻辑解析。
+
+---
+
+## 特殊标注值（realfoa 子集）
+
+`dcase_real/realfoa` 子集��自真实自录 FOA，存在两类标注缺失：
+
+### `distance_cm = -1`（距离未知）
+
+- 含义：该音源距离完全未知，无法估计。
+- **现有代码已正确处理**：`parse_csv_rows()` 执行 `distance_m = None if distance_cm < 0 else …`，后续所有距离 QA 任务均 gate 在 `span.distance_m is not None` 上，无需改动。
+
+### `elevation_deg = ±inf`（仰角符号已知，幅度未知）
+
+- `+inf`：确知音源在水平面**以上**，但仰角具体数值未知。
+- `-inf`：确知音源在水平面**以下**，但仰角具体数值未知。
+- **现有代码存在问题**：`float('inf')` 可无声地通过 `parse_csv_rows()`，但会导致：
+  - `finalize_event_span()` 中 `median([inf,...]) = inf` → `span.elevation_deg = inf`
+  - `format_angle(inf)` → `"inf degrees"`（垃圾答案）
+  - v3 elevation bin 分配失败（超出 [-90,90] 范围）
+  - `classify_motion()` 中 `elevation_delta = inf` → 错误标记"moving upward"
+  - `is_unique_doa_at_frame()` 中 `abs(inf - finite) = inf` → 判断错误
+- **处理方向（待实现，见待办）**：检测 `math.isinf(elevation_deg)` 后，跳过所有仰角幅度相关 QA，仅保留上/下方向（above/below horizon）判断类问题。
+
+---
+
+## 子集说明
+
+### `sim_static`（304,426 条）
+
+- 来源：基于 HM3D 房间的仿真 FOA，单源静态场景
+- 音源：FSD50K 音频 + 房间 RIR 卷积
+- 标注：完整（azimuth、elevation、distance 均有值）
+- 特点：方位角、仰角覆盖均匀，适合空间定位训练
+
+### `qa_sim`（74,891 条）
+
+- 来源：QA lr_pair 仿真场景（HM3D 房间，多源）
+- 标注：完整，含多源同时活跃的重叠场景
+- 特点：为 QA 训练特别构造，重叠事件比例较高
+
+### `dcase_real`（20,623 条）
+
+真实录音子集，包含以下 8 个来源：
+
+| 子集 | train | valid | test | 特点 |
+|------|-------|-------|------|------|
+| realfoa | 2,492 | 0 | 70 | 自录 FOA；**含 ±inf elevation、-1 distance** |
+| starss22 | 545 | 355 | 33 | Sony-TAu STARSS 2022 |
+| starss23 | 789 | 556 | 63 | Sony-TAu STARSS 2023 |
+| tau2019 | 994 | 292 | 34 | TAU 2019 空间音频 |
+| tau2020 | 1,050 | 675 | 75 | TAU 2020 |
+| tau2021 | 900 | 810 | 90 | TAU 2021 |
+| tau2636586 | 4,320 | 972 | 108 | TAU 合成扩充 |
+| tau2636594 | 4,320 | 972 | 108 | TAU 合成扩充 |
+
+---
+
+## 类别体系（FSD63）
+
+63 类，定义见 `label_mapping.json`，ID 0–62。
+
+常见高频类别（scene 数 top-10，来自 summary.json label_counts）：
+`speech (115k)`, `wind_instrument (25k)`, `human_vocalization (24k)`, `singing (24k)`, `tool (23k)`, `laughter (21k)`, `telephone_alarm (19k)`, `breathing (19k)`, `metal_clink (19k)`, `string_instrument (20k)`
+
+完整类别列表：
+```
+aircraft, alarm, animal, appliance, bell, bird, body_sound, breathing,
+camera, car, cat, clock, cooking, crack, crackle, crushing, dog, door,
+drawer_cabinet, drum, female_speech, finger_snapping, fire, footsteps,
+frog, glass, gong, guitar, home_sound, human_vocalization, insect,
+keyboard_instrument, kitchenware, knock, laughter, machine, male_speech,
+metal_clink, musical_instrument, ocean, paper, percussion, printer, rain,
+scratch, singing, speech, string_instrument, tape, tearing, telephone_alarm,
+thunderstorm, tool, train, typing, vehicle, war_sound, water, wind,
+wind_instrument, wood, writing, zipper
+```
+
+QA 生成时需通过 `--class-mapping-path label_mapping.json` 传入类别映射。
+
+---
+
+## 与已有训练流程的兼容性
+
+### 兼容项
+
+- CSV 格式与 DCASE 2024 基线完全一致（无 header，6 列，10 Hz）
+- `distance=-1` 已被 `parse_csv_rows()` 正确转为 `None`，所有距离 QA 已跳过
+- `float('inf')` 在 Python `csv.reader` + `float()` 下可正常解析，不会 crash
+- `scene_annotation_csv_path` 可直接作为 `csv_path` 传入 `parse_csv_rows()`
+
+### 不兼容项 / 待处理
+
+1. **JSONL 格式不兼容**：`load_segments()` 期望扁平 `{foa_path, csv_path, stem, length, ...}`，新数据为嵌套格式。需新增 `load_segments_unified()` 适配函数。
+
+2. **`elevation=±inf` 导致 QA 垃圾输出**：当前代码无 `math.isinf()` 保护，会生成 "inf degrees" 等无效答案，以及错误的运动检测结果。需在 QA 生成层加以处理（跳过 elevation 幅度 QA，仅保留上/下判断）。
+
+3. **类别 ID 空间扩展**：FSD63 的 63 个类别 ID（0–62）与旧有 DCASE 29 类不同，QA 生成时须正确传入 `label_mapping.json`。
+
+---
+
+## 待办事项
+
+- [ ] 在 `generate_segment_qa_pairs.py` 中新增 `load_segments_unified(dataset_root, limit_per_split)` 函数，解析新 JSONL schema 并输出 `SegmentRecord` 列表
+- [ ] 为 `elevation=±inf` 添加保护：`finalize_event_span()` 检测 `math.isinf(elevation_deg)`，为 `EventSpan` 增加 `elevation_sign_only: bool` 标志；`add_angle_records()`（v3）在 `elevation_sign_only=True` 时跳过 bin-choice 任务，可选生成新的 `above_below_horizon` 判断题
+- [ ] `classify_motion()` 中 `elevation_delta` 含 inf 时忽略垂直分量，避免误判"moving upward/downward"
+- [ ] `is_unique_doa_at_frame()` 中 elevation=inf 时跳过 elevation 比较维度
+- [ ] 用新数据生成 v6c QA pairs（建议单独输出目录，与旧 all_qa 合并后再做 v12 训练）
+- [ ] 开启 v12 训练，从 v11a stage1 `best_trainable.pt` 继承
+
+---
+
+## 参考
+
+- 数据路径：`/apdcephfs_cq12/share_302080740/user/schmittzhu/data/unified_spatial_foa_fsd63_all/`
+- 类别映射：同目录 `label_mapping.json`
+- 统计汇总：同目录 `summary.json`
+- QA 生成脚本：`DCASE2024_seld_baseline/generate_segment_qa_pairs_v6c.py`
+- 上游训练入口：`train_spatial_beats_qa.py`（BEATs 路径）

runs/qwen3_curriculum/bench_v4_via_train/easy.PRE_GEN_FIX_20260519_0014/epoch_metrics.jsonl

@@ -0,0 +1 @@
+{"epoch": 1, "train_loss": 0.664092443883419, "train_supervised_tokens": 168.0, "optimizer_steps": 1.0, "epoch_seconds": 2.743562936782837, "micro_steps": 1.0, "valid_loss": 0.7183872366560142, "valid_supervised_tokens": 69109.0, "valid_generate_examples": 4000.0, "valid_exact_match": 0.02575, "step_valid_subset_size": 200.0, "valid_subset_size": 4000.0, "valid_generation_size": 4000.0, "global_optimizer_step": 1.0}

runs/qwen3_curriculum/bench_v4_via_train/easy.PRE_GEN_FIX_20260519_0014/processor/processor_config.json

@@ -0,0 +1,35 @@
+{
+  "feature_extractor": {
+    "chunk_length": 30,
+    "dither": 0.0,
+    "feature_extractor_type": "WhisperFeatureExtractor",
+    "feature_size": 128,
+    "hop_length": 160,
+    "image_mean": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "image_processor_type": "Qwen2VLImageProcessor",
+    "image_std": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "max_pixels": 12845056,
+    "merge_size": 2,
+    "min_pixels": 3136,
+    "n_fft": 400,
+    "n_samples": 480000,
+    "nb_max_frames": 3000,
+    "padding_side": "right",
+    "padding_value": 0.0,
+    "patch_size": 16,
+    "return_attention_mask": true,
+    "sampling_rate": 16000,
+    "temporal_patch_size": 2
+  },
+  "image_processor": null,
+  "processor_class": "Qwen3OmniMoeSpatialProcessor",
+  "video_processor": null
+}

runs/qwen3_curriculum/bench_v4_via_train/easy.PRE_GEN_FIX_20260519_0014/processor/tokenizer_config.json

@@ -0,0 +1,34 @@
+{
+  "add_prefix_space": false,
+  "audio_bos_token": "<|audio_start|>",
+  "audio_eos_token": "<|audio_end|>",
+  "audio_token": "<|audio_pad|>",
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": [
+    "<|spatial|>"
+  ],
+  "image_token": "<|image_pad|>",
+  "is_local": true,
+  "model_max_length": 131072,
+  "model_specific_special_tokens": {
+    "audio_bos_token": "<|audio_start|>",
+    "audio_eos_token": "<|audio_end|>",
+    "audio_token": "<|audio_pad|>",
+    "image_token": "<|image_pad|>",
+    "video_token": "<|video_pad|>",
+    "vision_bos_token": "<|vision_start|>",
+    "vision_eos_token": "<|vision_end|>"
+  },
+  "pad_token": "<|endoftext|>",
+  "processor_class": "Qwen3OmniMoeSpatialProcessor",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null,
+  "video_token": "<|video_pad|>",
+  "vision_bos_token": "<|vision_start|>",
+  "vision_eos_token": "<|vision_end|>"
+}

runs/qwen3_curriculum/bench_v4_via_train/easy.PRE_GEN_FIX_20260519_0014/train_args.json

@@ -0,0 +1,75 @@
+{
+  "attn_impl": "flash_attention_2",
+  "audio_feature_cache_manifest": null,
+  "audio_feature_cache_max_entries": 256,
+  "batch_size": 1,
+  "beats_checkpoint": "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/checkpoints/spatial_beats_ov1_unified_v13d_exp/03_ov123_top4/best.pt",
+  "beats_lr": null,
+  "beats_repo": "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats",
+  "device": "cuda:0",
+  "device_map": null,
+  "distributed": true,
+  "dtype": "bfloat16",
+  "encoder_token_rate": 10.0,
+  "epochs": 1,
+  "grad_accum_steps": 1,
+  "gradient_checkpointing": true,
+  "local_rank": 0,
+  "log_every": 10,
+  "lora_alpha": 32,
+  "lora_dropout": 0.05,
+  "lora_lr": null,
+  "lora_r": 16,
+  "lora_target_modules": [
+    "q_proj",
+    "k_proj",
+    "v_proj",
+    "o_proj"
+  ],
+  "lora_target_prefixes": [
+    "model.layers"
+  ],
+  "lr": 0.0001,
+  "max_grad_norm": 1.0,
+  "max_train_samples": 1,
+  "max_valid_samples": null,
+  "model_id": "/tmp/Qwen3-Omni-30B-A3B-Instruct",
+  "num_workers": 4,
+  "optimizer_step_per_batch": false,
+  "output_dir": "runs/qwen3_curriculum/bench_v4_via_train/easy",
+  "persistent_workers": false,
+  "prefetch_factor": 2,
+  "projector_fp32": false,
+  "projector_lr": null,
+  "projector_shuffle_factor": 4,
+  "projector_type": "pixel_shuffle",
+  "projector_weight_decay": null,
+  "qa_root": "/apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2_filtered_balanced_v4_prompted/easy",
+  "qa_roots": [
+    "/apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2_filtered_balanced_v4_prompted/easy"
+  ],
+  "rank": 0,
+  "resume_checkpoint_path": "runs/qwen3_curriculum/easy/stage3_beats_lora/checkpoints/best_trainable.pt",
+  "resume_model_only": true,
+  "resume_tag": null,
+  "save_every_epoch": true,
+  "save_every_n_optimizer_steps": 1000,
+  "save_full_model": false,
+  "seed": 1234,
+  "spatial_qwen_repo": "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen",
+  "step_valid_subset_ratio": 0.05,
+  "train_mode": "beats_lora",
+  "train_split": "test",
+  "valid_do_sample": false,
+  "valid_every_n_optimizer_steps": 1000,
+  "valid_generate_batch_size": 1,
+  "valid_generate_full": true,
+  "valid_generate_max_samples": 32,
+  "valid_max_new_tokens": 80,
+  "valid_num_beams": 1,
+  "valid_split": "test",
+  "valid_subset_ratio": 0.1,
+  "warmup_ratio": 0.03,
+  "weight_decay": 0.01,
+  "world_size": 8
+}

runs/qwen3_curriculum/bench_v4_via_train/easy.PRE_GEN_FIX_20260519_0014/valid_predictions/epoch_001.score.json

@@ -0,0 +1,202 @@
+{
+  "examples": 4000,
+  "llm_used_rate": 0.0,
+  "overall_correct": 0.3130372081801082,
+  "overall_correct_parseable": 0.3194257226327635,
+  "parse_rate": 0.98,
+  "parse_status_counts": {
+    "fail_empty": 0,
+    "fail_llm_extract": 0,
+    "fail_no_answer_meta": 0,
+    "fail_regex": 80,
+    "ok": 3920
+  },
+  "per_task": {
+    "detect_source": {
+      "correct_all": 0.29439208966945096,
+      "correct_parseable": 0.29616821087409323,
+      "examples": 667,
+      "f1_mean": 0.29616821087409323,
+      "llm_used_rate": 0.0,
+      "metric_type": "detect_source_f1",
+      "parse_rate": 0.9940029985007496,
+      "parse_status_counts": {
+        "fail_empty": 0,
+        "fail_llm_extract": 0,
+        "fail_no_answer_meta": 0,
+        "fail_regex": 4,
+        "ok": 663
+      },
+      "precision_mean": 0.33056812468577174,
+      "recall_mean": 0.28179989944695827
+    },
+    "detect_time": {
+      "correct_all": 0.5529075096115579,
+      "correct_parseable": 0.5554055857092005,
+      "end_error_mean_s": 2.6058618222891567,
+      "examples": 667,
+      "iou_at_threshold": 0.6024096385542169,
+      "iou_mean": 0.5554055857092005,
+      "iou_median": 0.6086956521739131,
+      "llm_used_rate": 0.0,
+      "metric_type": "time_span_iou",
+      "parse_rate": 0.9955022488755623,
+      "parse_status_counts": {
+        "fail_empty": 0,
+        "fail_llm_extract": 0,
+        "fail_no_answer_meta": 0,
+        "fail_regex": 3,
+        "ok": 664
+      },
+      "start_error_mean_s": 2.066867469879518
+    },
+    "estimate_azimuth": {
+      "acc_gain_over_constant": 0.005641519381977406,
+      "acc_macro_by_gt_bin": 0.14531420552003743,
+      "acc_per_gt_bin": {
+        "[-180,-90)": {
+          "acc": 0.0,
+          "n": 173
+        },
+        "[-30,30)": {
+          "acc": 0.5943396226415094,
+          "n": 106
+        },
+        "[-90,-30)": {
+          "acc": 0.1322314049586777,
+          "n": 121
+        },
+        "[30,90)": {
+          "acc": 0.0,
+          "n": 100
+        },
+        "[90,180)": {
+          "acc": 0.0,
+          "n": 155
+        }
+      },
+      "const_median_baseline_acc": 0.11297709923664122,
+      "const_median_baseline_mean_err": 89.39981908396948,
+      "const_median_baseline_median_err": 90.09300000000002,
+      "const_median_baseline_value_deg": -10.711,
+      "correct_all": 0.11861861861861862,
+      "correct_is_within_threshold": 0.11861861861861862,
+      "correct_parseable": 0.12061068702290076,
+      "error_deg_mean": 87.53453816793893,
+      "error_deg_median": 87.72200000000001,
+      "examples": 666,
+      "llm_used_rate": 0.0,
+      "median_err_reduction_vs_constant": 2.3710000000000093,
+      "metric_type": "er20_angle",
+      "parse_rate": 0.9834834834834835,
+      "parse_status_counts": {
+        "fail_empty": 0,
+        "fail_llm_extract": 0,
+        "fail_no_answer_meta": 0,
+        "fail_regex": 11,
+        "ok": 655
+      },
+      "top1_prediction": -18.1,
+      "top1_prediction_share": 0.1648854961832061,
+      "unique_prediction_count": 88,
+      "unique_prediction_ratio": 0.13435114503816795
+    },
+    "estimate_elevation": {
+      "acc_gain_over_constant": -0.04618186774381405,
+      "acc_macro_by_gt_bin": 0.22468907305785274,
+      "acc_per_gt_bin": {
+        "[-10,10)": {
+          "acc": 0.35960591133004927,
+          "n": 203
+        },
+        "[-30,-10)": {
+          "acc": 0.6467065868263473,
+          "n": 167
+        },
+        "[-90,-30)": {
+          "acc": 0.011363636363636364,
+          "n": 88
+        },
+        "[10,30)": {
+          "acc": 0.10576923076923077,
+          "n": 104
+        },
+        "[30,90)": {
+          "acc": 0.0,
+          "n": 42
+        }
+      },
+      "const_median_baseline_acc": 0.33553719008264465,
+      "const_median_baseline_mean_err": Infinity,
+      "const_median_baseline_median_err": 15.631,
+      "const_median_baseline_value_deg": -3.45,
+      "correct_all": 0.2893553223388306,
+      "correct_is_within_threshold": 0.2893553223388306,
+      "correct_parseable": 0.31900826446280994,
+      "error_deg_mean": Infinity,
+      "error_deg_median": 16.734,
+      "examples": 667,
+      "llm_used_rate": 0.0,
+      "median_err_reduction_vs_constant": -1.1030000000000015,
+      "metric_type": "abs10_angle",
+      "parse_rate": 0.9070464767616192,
+      "parse_status_counts": {
+        "fail_empty": 0,
+        "fail_llm_extract": 0,
+        "fail_no_answer_meta": 0,
+        "fail_regex": 62,
+        "ok": 605
+      },
+      "top1_prediction": -18.1,
+      "top1_prediction_share": 0.24462809917355371,
+      "unique_prediction_count": 68,
+      "unique_prediction_ratio": 0.11239669421487604
+    },
+    "identify_source_by_doa": {
+      "correct_all": 0.2158920539730135,
+      "correct_parseable": 0.2158920539730135,
+      "examples": 667,
+      "llm_used_rate": 0.0,
+      "match_stage_counts": {
+        "none": 523,
+        "substring": 144
+      },
+      "metric_type": "source_label_match",
+      "parse_rate": 1.0,
+      "parse_status_counts": {
+        "fail_empty": 0,
+        "fail_llm_extract": 0,
+        "fail_no_answer_meta": 0,
+        "fail_regex": 0,
+        "ok": 667
+      }
+    },
+    "identify_source_by_location": {
+      "correct_all": 0.4069069069069069,
+      "correct_parseable": 0.4069069069069069,
+      "examples": 666,
+      "llm_used_rate": 0.0,
+      "match_stage_counts": {
+        "none": 395,
+        "substring": 271
+      },
+      "metric_type": "source_label_match",
+      "parse_rate": 1.0,
+      "parse_status_counts": {
+        "fail_empty": 0,
+        "fail_llm_extract": 0,
+        "fail_no_answer_meta": 0,
+        "fail_regex": 0,
+        "ok": 666
+      }
+    }
+  },
+  "thresholds": {
+    "angle_deg": 20.0,
+    "azimuth_deg": 20.0,
+    "distance_rel": 0.3,
+    "elevation_deg": 10.0,
+    "iou": 0.5,
+    "onset_s": 0.4
+  }
+}

runs/qwen3_curriculum/easy/stage3_beats_lora/bench_v4_via_train/easy/epoch_metrics.jsonl

@@ -0,0 +1 @@
+{"epoch": 1, "train_loss": 0.7423601821064949, "train_supervised_tokens": 168.0, "optimizer_steps": 1.0, "epoch_seconds": 2.3404576778411865, "micro_steps": 1.0, "valid_loss": 0.719199836174988, "valid_supervised_tokens": 69109.0, "valid_generate_examples": 4000.0, "valid_exact_match": 0.02425, "step_valid_subset_size": 200.0, "valid_subset_size": 4000.0, "valid_generation_size": 4000.0, "global_optimizer_step": 1.0}

runs/qwen3_curriculum/easy/stage3_beats_lora/bench_v4_via_train/easy/processor/processor_config.json

@@ -0,0 +1,35 @@
+{
+  "feature_extractor": {
+    "chunk_length": 30,
+    "dither": 0.0,
+    "feature_extractor_type": "WhisperFeatureExtractor",
+    "feature_size": 128,
+    "hop_length": 160,
+    "image_mean": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "image_processor_type": "Qwen2VLImageProcessor",
+    "image_std": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "max_pixels": 12845056,
+    "merge_size": 2,
+    "min_pixels": 3136,
+    "n_fft": 400,
+    "n_samples": 480000,
+    "nb_max_frames": 3000,
+    "padding_side": "right",
+    "padding_value": 0.0,
+    "patch_size": 16,
+    "return_attention_mask": true,
+    "sampling_rate": 16000,
+    "temporal_patch_size": 2
+  },
+  "image_processor": null,
+  "processor_class": "Qwen3OmniMoeSpatialProcessor",
+  "video_processor": null
+}

runs/qwen3_curriculum/easy/stage3_beats_lora/bench_v4_via_train/easy/processor/tokenizer_config.json

@@ -0,0 +1,34 @@
+{
+  "add_prefix_space": false,
+  "audio_bos_token": "<|audio_start|>",
+  "audio_eos_token": "<|audio_end|>",
+  "audio_token": "<|audio_pad|>",
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": [
+    "<|spatial|>"
+  ],
+  "image_token": "<|image_pad|>",
+  "is_local": true,
+  "model_max_length": 131072,
+  "model_specific_special_tokens": {
+    "audio_bos_token": "<|audio_start|>",
+    "audio_eos_token": "<|audio_end|>",
+    "audio_token": "<|audio_pad|>",
+    "image_token": "<|image_pad|>",
+    "video_token": "<|video_pad|>",
+    "vision_bos_token": "<|vision_start|>",
+    "vision_eos_token": "<|vision_end|>"
+  },
+  "pad_token": "<|endoftext|>",
+  "processor_class": "Qwen3OmniMoeSpatialProcessor",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null,
+  "video_token": "<|video_pad|>",
+  "vision_bos_token": "<|vision_start|>",
+  "vision_eos_token": "<|vision_end|>"
+}

runs/qwen3_curriculum/easy/stage3_beats_lora/bench_v4_via_train/easy/train_args.json

@@ -0,0 +1,75 @@
+{
+  "attn_impl": "flash_attention_2",
+  "audio_feature_cache_manifest": null,
+  "audio_feature_cache_max_entries": 256,
+  "batch_size": 1,
+  "beats_checkpoint": "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/checkpoints/spatial_beats_ov1_unified_v13d_exp/03_ov123_top4/best.pt",
+  "beats_lr": null,
+  "beats_repo": "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats",
+  "device": "cuda:0",
+  "device_map": null,
+  "distributed": true,
+  "dtype": "bfloat16",
+  "encoder_token_rate": 10.0,
+  "epochs": 1,
+  "grad_accum_steps": 1,
+  "gradient_checkpointing": true,
+  "local_rank": 0,
+  "log_every": 10,
+  "lora_alpha": 32,
+  "lora_dropout": 0.05,
+  "lora_lr": null,
+  "lora_r": 16,
+  "lora_target_modules": [
+    "q_proj",
+    "k_proj",
+    "v_proj",
+    "o_proj"
+  ],
+  "lora_target_prefixes": [
+    "model.layers"
+  ],
+  "lr": 0.0001,
+  "max_grad_norm": 1.0,
+  "max_train_samples": 1,
+  "max_valid_samples": null,
+  "model_id": "/tmp/Qwen3-Omni-30B-A3B-Instruct",
+  "num_workers": 4,
+  "optimizer_step_per_batch": false,
+  "output_dir": "runs/qwen3_curriculum/easy/stage3_beats_lora/bench_v4_via_train/easy",
+  "persistent_workers": true,
+  "prefetch_factor": 2,
+  "projector_fp32": false,
+  "projector_lr": null,
+  "projector_shuffle_factor": 4,
+  "projector_type": "pixel_shuffle",
+  "projector_weight_decay": null,
+  "qa_root": "/apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2_filtered_balanced_v4_prompted/easy",
+  "qa_roots": [
+    "/apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2_filtered_balanced_v4_prompted/easy"
+  ],
+  "rank": 0,
+  "resume_checkpoint_path": "runs/qwen3_curriculum/easy/stage3_beats_lora/checkpoints/best_trainable.pt",
+  "resume_model_only": true,
+  "resume_tag": null,
+  "save_every_epoch": true,
+  "save_every_n_optimizer_steps": 999999,
+  "save_full_model": false,
+  "seed": 1234,
+  "spatial_qwen_repo": "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen",
+  "step_valid_subset_ratio": 0.05,
+  "train_mode": "beats_lora",
+  "train_split": "test",
+  "valid_do_sample": false,
+  "valid_every_n_optimizer_steps": 999999,
+  "valid_generate_batch_size": 1,
+  "valid_generate_full": true,
+  "valid_generate_max_samples": 32,
+  "valid_max_new_tokens": 80,
+  "valid_num_beams": 1,
+  "valid_split": "test",
+  "valid_subset_ratio": 0.1,
+  "warmup_ratio": 0.03,
+  "weight_decay": 0.01,
+  "world_size": 8
+}

runs/qwen3_curriculum/easy/stage3_beats_lora/bench_v4_via_train/easy/valid_predictions/epoch_001.score.json

@@ -0,0 +1,202 @@
+{
+  "examples": 4000,
+  "llm_used_rate": 0.0,
+  "overall_correct": 0.3090666965864644,
+  "overall_correct_parseable": 0.316019117164074,
+  "parse_rate": 0.978,
+  "parse_status_counts": {
+    "fail_empty": 0,
+    "fail_llm_extract": 0,
+    "fail_no_answer_meta": 0,
+    "fail_regex": 88,
+    "ok": 3912
+  },
+  "per_task": {
+    "detect_source": {
+      "correct_all": 0.2994038694938245,
+      "correct_parseable": 0.30303851434352197,
+      "examples": 667,
+      "f1_mean": 0.30303851434352197,
+      "llm_used_rate": 0.0,
+      "metric_type": "detect_source_f1",
+      "parse_rate": 0.9880059970014993,
+      "parse_status_counts": {
+        "fail_empty": 0,
+        "fail_llm_extract": 0,
+        "fail_no_answer_meta": 0,
+        "fail_regex": 8,
+        "ok": 659
+      },
+      "precision_mean": 0.34673748103186647,
+      "recall_mean": 0.2862923621648963
+    },
+    "detect_time": {
+      "correct_all": 0.5480725718043127,
+      "correct_parseable": 0.5522120927393905,
+      "end_error_mean_s": 2.5945502265861027,
+      "examples": 667,
+      "iou_at_threshold": 0.6057401812688822,
+      "iou_mean": 0.5522120927393905,
+      "iou_median": 0.606203007518797,
+      "llm_used_rate": 0.0,
+      "metric_type": "time_span_iou",
+      "parse_rate": 0.992503748125937,
+      "parse_status_counts": {
+        "fail_empty": 0,
+        "fail_llm_extract": 0,
+        "fail_no_answer_meta": 0,
+        "fail_regex": 5,
+        "ok": 662
+      },
+      "start_error_mean_s": 2.058308157099698
+    },
+    "estimate_azimuth": {
+      "acc_gain_over_constant": 0.0009918175055789735,
+      "acc_macro_by_gt_bin": 0.13834095673718314,
+      "acc_per_gt_bin": {
+        "[-180,-90)": {
+          "acc": 0.0,
+          "n": 174
+        },
+        "[-30,30)": {
+          "acc": 0.5566037735849056,
+          "n": 106
+        },
+        "[-90,-30)": {
+          "acc": 0.125,
+          "n": 120
+        },
+        "[30,90)": {
+          "acc": 0.010101010101010102,
+          "n": 99
+        },
+        "[90,180)": {
+          "acc": 0.0,
+          "n": 155
+        }
+      },
+      "const_median_baseline_acc": 0.11162079510703364,
+      "const_median_baseline_mean_err": 89.5508264525994,
+      "const_median_baseline_median_err": 90.50300000000001,
+      "const_median_baseline_value_deg": -11.294,
+      "correct_all": 0.11261261261261261,
+      "correct_is_within_threshold": 0.11261261261261261,
+      "correct_parseable": 0.11467889908256881,
+      "error_deg_mean": 87.97856039755351,
+      "error_deg_median": 88.28200000000001,
+      "examples": 666,
+      "llm_used_rate": 0.0,
+      "median_err_reduction_vs_constant": 2.2210000000000036,
+      "metric_type": "er20_angle",
+      "parse_rate": 0.9819819819819819,
+      "parse_status_counts": {
+        "fail_empty": 0,
+        "fail_llm_extract": 0,
+        "fail_no_answer_meta": 0,
+        "fail_regex": 12,
+        "ok": 654
+      },
+      "top1_prediction": -18.1,
+      "top1_prediction_share": 0.16055045871559634,
+      "unique_prediction_count": 88,
+      "unique_prediction_ratio": 0.1345565749235474
+    },
+    "estimate_elevation": {
+      "acc_gain_over_constant": -0.0476111282107291,
+      "acc_macro_by_gt_bin": 0.22207134493831174,
+      "acc_per_gt_bin": {
+        "[-10,10)": {
+          "acc": 0.3382352941176471,
+          "n": 204
+        },
+        "[-30,-10)": {
+          "acc": 0.6463414634146342,
+          "n": 164
+        },
+        "[-90,-30)": {
+          "acc": 0.011494252873563218,
+          "n": 87
+        },
+        "[10,30)": {
+          "acc": 0.11428571428571428,
+          "n": 105
+        },
+        "[30,90)": {
+          "acc": 0.0,
+          "n": 43
+        }
+      },
+      "const_median_baseline_acc": 0.3294701986754967,
+      "const_median_baseline_mean_err": Infinity,
+      "const_median_baseline_median_err": 15.6485,
+      "const_median_baseline_value_deg": -2.452,
+      "correct_all": 0.2818590704647676,
+      "correct_is_within_threshold": 0.2818590704647676,
+      "correct_parseable": 0.31125827814569534,
+      "error_deg_mean": Infinity,
+      "error_deg_median": 17.038,
+      "examples": 667,
+      "llm_used_rate": 0.0,
+      "median_err_reduction_vs_constant": -1.3895,
+      "metric_type": "abs10_angle",
+      "parse_rate": 0.9055472263868066,
+      "parse_status_counts": {
+        "fail_empty": 0,
+        "fail_llm_extract": 0,
+        "fail_no_answer_meta": 0,
+        "fail_regex": 63,
+        "ok": 604
+      },
+      "top1_prediction": -18.1,
+      "top1_prediction_share": 0.23344370860927152,
+      "unique_prediction_count": 70,
+      "unique_prediction_ratio": 0.11589403973509933
+    },
+    "identify_source_by_doa": {
+      "correct_all": 0.21739130434782608,
+      "correct_parseable": 0.21739130434782608,
+      "examples": 667,
+      "llm_used_rate": 0.0,
+      "match_stage_counts": {
+        "none": 522,
+        "substring": 145
+      },
+      "metric_type": "source_label_match",
+      "parse_rate": 1.0,
+      "parse_status_counts": {
+        "fail_empty": 0,
+        "fail_llm_extract": 0,
+        "fail_no_answer_meta": 0,
+        "fail_regex": 0,
+        "ok": 667
+      }
+    },
+    "identify_source_by_location": {
+      "correct_all": 0.3948948948948949,
+      "correct_parseable": 0.3948948948948949,
+      "examples": 666,
+      "llm_used_rate": 0.0,
+      "match_stage_counts": {
+        "none": 403,
+        "substring": 263
+      },
+      "metric_type": "source_label_match",
+      "parse_rate": 1.0,
+      "parse_status_counts": {
+        "fail_empty": 0,
+        "fail_llm_extract": 0,
+        "fail_no_answer_meta": 0,
+        "fail_regex": 0,
+        "ok": 666
+      }
+    }
+  },
+  "thresholds": {
+    "angle_deg": 20.0,
+    "azimuth_deg": 20.0,
+    "distance_rel": 0.3,
+    "elevation_deg": 10.0,
+    "iou": 0.5,
+    "onset_s": 0.4
+  }
+}

runs/qwen3_curriculum/easy/stage3_beats_lora/epoch_metrics.jsonl

@@ -0,0 +1,3 @@
+{"epoch": 1, "train_loss": 0.3592844999662934, "train_supervised_tokens": 11843369.0, "optimizer_steps": 9796.0, "epoch_seconds": 51669.55358290672, "micro_steps": 39181.0, "valid_loss": 0.4098466007066923, "valid_supervised_tokens": 1289631.0, "valid_generate_examples": 32.0, "valid_exact_match": 0.0, "step_valid_subset_size": 3292.0, "valid_subset_size": 65850.0, "valid_generation_size": 32.0, "global_optimizer_step": 9796.0}
+{"epoch": 2, "train_loss": 0.3392781011698098, "train_supervised_tokens": 11843369.0, "optimizer_steps": 9796.0, "epoch_seconds": 53220.040230989456, "micro_steps": 39181.0, "valid_loss": 0.39830992954270594, "valid_supervised_tokens": 1289631.0, "valid_generate_examples": 32.0, "valid_exact_match": 0.0, "step_valid_subset_size": 3292.0, "valid_subset_size": 65850.0, "valid_generation_size": 32.0, "global_optimizer_step": 19592.0}
+{"epoch": 3, "train_loss": 0.32576761328400655, "train_supervised_tokens": 11843369.0, "optimizer_steps": 9796.0, "epoch_seconds": 54061.63148808479, "micro_steps": 39181.0, "valid_loss": 0.3947276146043098, "valid_supervised_tokens": 1289631.0, "valid_generate_examples": 32.0, "valid_exact_match": 0.0, "step_valid_subset_size": 3292.0, "valid_subset_size": 65850.0, "valid_generation_size": 32.0, "global_optimizer_step": 29388.0}

runs/qwen3_curriculum/easy/stage3_beats_lora/train_args.json

@@ -0,0 +1,75 @@
+{
+  "attn_impl": "flash_attention_2",
+  "audio_feature_cache_manifest": null,
+  "audio_feature_cache_max_entries": 256,
+  "batch_size": 2,
+  "beats_checkpoint": "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/checkpoints/spatial_beats_ov1_unified_v13d_exp/03_ov123_top4/best.pt",
+  "beats_lr": 1e-06,
+  "beats_repo": "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats",
+  "device": "cuda:0",
+  "device_map": null,
+  "distributed": true,
+  "dtype": "bfloat16",
+  "encoder_token_rate": 10.0,
+  "epochs": 3,
+  "grad_accum_steps": 4,
+  "gradient_checkpointing": true,
+  "local_rank": 0,
+  "log_every": 10,
+  "lora_alpha": 32,
+  "lora_dropout": 0.05,
+  "lora_lr": 3e-05,
+  "lora_r": 16,
+  "lora_target_modules": [
+    "q_proj",
+    "k_proj",
+    "v_proj",
+    "o_proj"
+  ],
+  "lora_target_prefixes": [
+    "model.layers"
+  ],
+  "lr": 3e-05,
+  "max_grad_norm": 1.0,
+  "max_train_samples": null,
+  "max_valid_samples": null,
+  "model_id": "/tmp/Qwen3-Omni-30B-A3B-Instruct",
+  "num_workers": 4,
+  "optimizer_step_per_batch": false,
+  "output_dir": "./runs/qwen3_curriculum/easy/stage3_beats_lora",
+  "persistent_workers": true,
+  "prefetch_factor": 2,
+  "projector_fp32": false,
+  "projector_lr": 1e-06,
+  "projector_shuffle_factor": 4,
+  "projector_type": "pixel_shuffle",
+  "projector_weight_decay": null,
+  "qa_root": "/apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2_filtered_balanced/easy",
+  "qa_roots": [
+    "/apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2_filtered_balanced/easy"
+  ],
+  "rank": 0,
+  "resume_checkpoint_path": "./runs/qwen3_curriculum/easy/stage2_encoder_lora/checkpoints/best_trainable.pt",
+  "resume_model_only": true,
+  "resume_tag": null,
+  "save_every_epoch": true,
+  "save_every_n_optimizer_steps": 500,
+  "save_full_model": false,
+  "seed": 1234,
+  "spatial_qwen_repo": "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen",
+  "step_valid_subset_ratio": 0.05,
+  "train_mode": "beats_lora",
+  "train_split": "train",
+  "valid_do_sample": false,
+  "valid_every_n_optimizer_steps": 500,
+  "valid_generate_batch_size": 1,
+  "valid_generate_full": false,
+  "valid_generate_max_samples": 32,
+  "valid_max_new_tokens": 96,
+  "valid_num_beams": 1,
+  "valid_split": "valid",
+  "valid_subset_ratio": 0.1,
+  "warmup_ratio": 0.03,
+  "weight_decay": 0.01,
+  "world_size": 8
+}

runs/qwen3_full_v4_easy/stage_encoder_lora/epoch_metrics.jsonl

@@ -0,0 +1 @@
+{"epoch": 1, "train_loss": 0.4386036706659937, "train_supervised_tokens": 11843369.0, "optimizer_steps": 9796.0, "epoch_seconds": 43788.69994306564, "micro_steps": 39181.0, "valid_loss": 0.4266115554512279, "valid_supervised_tokens": 69109.0, "valid_generate_examples": 4000.0, "valid_exact_match": 0.1125, "step_valid_subset_size": 1.0, "valid_subset_size": 4000.0, "valid_generation_size": 4000.0, "global_optimizer_step": 9796.0}

runs/qwen3_full_v4_easy/stage_encoder_lora/processor/processor_config.json

@@ -0,0 +1,35 @@
+{
+  "feature_extractor": {
+    "chunk_length": 30,
+    "dither": 0.0,
+    "feature_extractor_type": "WhisperFeatureExtractor",
+    "feature_size": 128,
+    "hop_length": 160,
+    "image_mean": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "image_processor_type": "Qwen2VLImageProcessor",
+    "image_std": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "max_pixels": 12845056,
+    "merge_size": 2,
+    "min_pixels": 3136,
+    "n_fft": 400,
+    "n_samples": 480000,
+    "nb_max_frames": 3000,
+    "padding_side": "right",
+    "padding_value": 0.0,
+    "patch_size": 16,
+    "return_attention_mask": true,
+    "sampling_rate": 16000,
+    "temporal_patch_size": 2
+  },
+  "image_processor": null,
+  "processor_class": "Qwen3OmniMoeSpatialProcessor",
+  "video_processor": null
+}

runs/qwen3_full_v4_easy/stage_encoder_lora/processor/tokenizer_config.json

@@ -0,0 +1,34 @@
+{
+  "add_prefix_space": false,
+  "audio_bos_token": "<|audio_start|>",
+  "audio_eos_token": "<|audio_end|>",
+  "audio_token": "<|audio_pad|>",
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": [
+    "<|spatial|>"
+  ],
+  "image_token": "<|image_pad|>",
+  "is_local": true,
+  "model_max_length": 131072,
+  "model_specific_special_tokens": {
+    "audio_bos_token": "<|audio_start|>",
+    "audio_eos_token": "<|audio_end|>",
+    "audio_token": "<|audio_pad|>",
+    "image_token": "<|image_pad|>",
+    "video_token": "<|video_pad|>",
+    "vision_bos_token": "<|vision_start|>",
+    "vision_eos_token": "<|vision_end|>"
+  },
+  "pad_token": "<|endoftext|>",
+  "processor_class": "Qwen3OmniMoeSpatialProcessor",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null,
+  "video_token": "<|video_pad|>",
+  "vision_bos_token": "<|vision_start|>",
+  "vision_eos_token": "<|vision_end|>"
+}

runs/qwen3_full_v4_easy/stage_encoder_lora/train_args.json

@@ -0,0 +1,75 @@
+{
+  "attn_impl": "flash_attention_2",
+  "audio_feature_cache_manifest": null,
+  "audio_feature_cache_max_entries": 256,
+  "batch_size": 2,
+  "beats_checkpoint": "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/checkpoints/spatial_beats_ov1_unified_v13d_exp/03_ov123_top4/best.pt",
+  "beats_lr": null,
+  "beats_repo": "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats",
+  "device": "cuda:0",
+  "device_map": null,
+  "distributed": true,
+  "dtype": "bfloat16",
+  "encoder_token_rate": 10.0,
+  "epochs": 2,
+  "grad_accum_steps": 4,
+  "gradient_checkpointing": true,
+  "local_rank": 0,
+  "log_every": 10,
+  "lora_alpha": 32,
+  "lora_dropout": 0.05,
+  "lora_lr": 5e-05,
+  "lora_r": 16,
+  "lora_target_modules": [
+    "q_proj",
+    "k_proj",
+    "v_proj",
+    "o_proj"
+  ],
+  "lora_target_prefixes": [
+    "model.layers"
+  ],
+  "lr": 5e-05,
+  "max_grad_norm": 1.0,
+  "max_train_samples": null,
+  "max_valid_samples": null,
+  "model_id": "/tmp/Qwen3-Omni-30B-A3B-Instruct",
+  "num_workers": 4,
+  "optimizer_step_per_batch": false,
+  "output_dir": "/apdcephfs_fsgm/share_303840540/hunyuan/jensenwang/zzy/Spatial-Qwen/runs/qwen3_full_v4_easy/stage_encoder_lora",
+  "persistent_workers": true,
+  "prefetch_factor": 2,
+  "projector_fp32": false,
+  "projector_lr": 3e-05,
+  "projector_shuffle_factor": 4,
+  "projector_type": "pixel_shuffle",
+  "projector_weight_decay": null,
+  "qa_root": "/apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2_filtered_balanced_v4/easy",
+  "qa_roots": [
+    "/apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2_filtered_balanced_v4/easy"
+  ],
+  "rank": 0,
+  "resume_checkpoint_path": "/apdcephfs_fsgm/share_303840540/hunyuan/jensenwang/zzy/Spatial-Qwen/runs/qwen3_quick_v4_easy_10pct/stage_encoder_lora_ep2-3/checkpoints/best_trainable.pt",
+  "resume_model_only": true,
+  "resume_tag": null,
+  "save_every_epoch": true,
+  "save_every_n_optimizer_steps": 5000,
+  "save_full_model": false,
+  "seed": 1234,
+  "spatial_qwen_repo": "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen",
+  "step_valid_subset_ratio": 0.0,
+  "train_mode": "encoder_lora",
+  "train_split": "train",
+  "valid_do_sample": false,
+  "valid_every_n_optimizer_steps": 999999,
+  "valid_generate_batch_size": 1,
+  "valid_generate_full": true,
+  "valid_generate_max_samples": 32,
+  "valid_max_new_tokens": 96,
+  "valid_num_beams": 1,
+  "valid_split": "test",
+  "valid_subset_ratio": 1.0,
+  "warmup_ratio": 0.03,
+  "weight_decay": 0.01,
+  "world_size": 8
+}

runs/qwen3_full_v4_easy/stage_encoder_lora/valid_predictions/epoch_001.score.json

@@ -0,0 +1,202 @@
+{
+  "examples": 4000,
+  "llm_used_rate": 0.0,
+  "overall_correct": 0.5682040188178945,
+  "overall_correct_parseable": 0.5682040188178945,
+  "parse_rate": 1.0,
+  "parse_status_counts": {
+    "fail_empty": 0,
+    "fail_llm_extract": 0,
+    "fail_no_answer_meta": 0,
+    "fail_regex": 0,
+    "ok": 4000
+  },
+  "per_task": {
+    "detect_source": {
+      "correct_all": 0.2994716927250661,
+      "correct_parseable": 0.2994716927250661,
+      "examples": 667,
+      "f1_mean": 0.2994716927250661,
+      "llm_used_rate": 0.0,
+      "metric_type": "detect_source_f1",
+      "parse_rate": 1.0,
+      "parse_status_counts": {
+        "fail_empty": 0,
+        "fail_llm_extract": 0,
+        "fail_no_answer_meta": 0,
+        "fail_regex": 0,
+        "ok": 667
+      },
+      "precision_mean": 0.3038230884557721,
+      "recall_mean": 0.2975655029628043
+    },
+    "detect_time": {
+      "correct_all": 0.703251058806535,
+      "correct_parseable": 0.703251058806535,
+      "end_error_mean_s": 1.5427623688155923,
+      "examples": 667,
+      "iou_at_threshold": 0.7436281859070465,
+      "iou_mean": 0.703251058806535,
+      "iou_median": 0.8749999999999999,
+      "llm_used_rate": 0.0,
+      "metric_type": "time_span_iou",
+      "parse_rate": 1.0,
+      "parse_status_counts": {
+        "fail_empty": 0,
+        "fail_llm_extract": 0,
+        "fail_no_answer_meta": 0,
+        "fail_regex": 0,
+        "ok": 667
+      },
+      "start_error_mean_s": 1.9649175412293851
+    },
+    "estimate_azimuth": {
+      "acc_gain_over_constant": 0.49549549549549554,
+      "acc_macro_by_gt_bin": 0.6013221746381812,
+      "acc_per_gt_bin": {
+        "[-180,-90)": {
+          "acc": 0.7068965517241379,
+          "n": 174
+        },
+        "[-30,30)": {
+          "acc": 0.6542056074766355,
+          "n": 107
+        },
+        "[-90,-30)": {
+          "acc": 0.5983606557377049,
+          "n": 122
+        },
+        "[30,90)": {
+          "acc": 0.5533980582524272,
+          "n": 103
+        },
+        "[90,180)": {
+          "acc": 0.49375,
+          "n": 160
+        }
+      },
+      "const_median_baseline_acc": 0.10810810810810811,
+      "const_median_baseline_mean_err": 89.60612087087087,
+      "const_median_baseline_median_err": 91.30099999999999,
+      "const_median_baseline_value_deg": -9.4565,
+      "correct_all": 0.6036036036036037,
+      "correct_is_within_threshold": 0.6036036036036037,
+      "correct_parseable": 0.6036036036036037,
+      "error_deg_mean": 31.44257282282282,
+      "error_deg_median": 14.547999999999996,
+      "examples": 666,
+      "llm_used_rate": 0.0,
+      "median_err_reduction_vs_constant": 76.75299999999999,
+      "metric_type": "er20_angle",
+      "parse_rate": 1.0,
+      "parse_status_counts": {
+        "fail_empty": 0,
+        "fail_llm_extract": 0,
+        "fail_no_answer_meta": 0,
+        "fail_regex": 0,
+        "ok": 666
+      },
+      "top1_prediction": -144.2,
+      "top1_prediction_share": 0.036036036036036036,
+      "unique_prediction_count": 250,
+      "unique_prediction_ratio": 0.37537537537537535
+    },
+    "estimate_elevation": {
+      "acc_gain_over_constant": 0.3148425787106447,
+      "acc_macro_by_gt_bin": 0.567338611760575,
+      "acc_per_gt_bin": {
+        "[-10,10)": {
+          "acc": 0.7022222222222222,
+          "n": 225
+        },
+        "[-30,-10)": {
+          "acc": 0.7821229050279329,
+          "n": 179
+        },
+        "[-90,-30)": {
+          "acc": 0.422680412371134,
+          "n": 97
+        },
+        "[10,30)": {
+          "acc": 0.6470588235294118,
+          "n": 119
+        },
+        "[30,90)": {
+          "acc": 0.2826086956521739,
+          "n": 46
+        }
+      },
+      "const_median_baseline_acc": 0.328335832083958,
+      "const_median_baseline_mean_err": Infinity,
+      "const_median_baseline_median_err": 15.456999999999999,
+      "const_median_baseline_value_deg": -2.844,
+      "correct_all": 0.6431784107946027,
+      "correct_is_within_threshold": 0.6431784107946027,
+      "correct_parseable": 0.6431784107946027,
+      "error_deg_mean": Infinity,
+      "error_deg_median": 6.688999999999999,
+      "examples": 667,
+      "llm_used_rate": 0.0,
+      "median_err_reduction_vs_constant": 8.768,
+      "metric_type": "abs10_angle",
+      "parse_rate": 1.0,
+      "parse_status_counts": {
+        "fail_empty": 0,
+        "fail_llm_extract": 0,
+        "fail_no_answer_meta": 0,
+        "fail_regex": 0,
+        "ok": 667
+      },
+      "top1_prediction": -10.0,
+      "top1_prediction_share": 0.10944527736131934,
+      "unique_prediction_count": 83,
+      "unique_prediction_ratio": 0.12443778110944528
+    },
+    "identify_source_by_doa": {
+      "correct_all": 0.6101949025487257,
+      "correct_parseable": 0.6101949025487257,
+      "examples": 667,
+      "llm_used_rate": 0.0,
+      "match_stage_counts": {
+        "none": 260,
+        "substring": 407
+      },
+      "metric_type": "source_label_match",
+      "parse_rate": 1.0,
+      "parse_status_counts": {
+        "fail_empty": 0,
+        "fail_llm_extract": 0,
+        "fail_no_answer_meta": 0,
+        "fail_regex": 0,
+        "ok": 667
+      }
+    },
+    "identify_source_by_location": {
+      "correct_all": 0.5495495495495496,
+      "correct_parseable": 0.5495495495495496,
+      "examples": 666,
+      "llm_used_rate": 0.0,
+      "match_stage_counts": {
+        "none": 300,
+        "substring": 366
+      },
+      "metric_type": "source_label_match",
+      "parse_rate": 1.0,
+      "parse_status_counts": {
+        "fail_empty": 0,
+        "fail_llm_extract": 0,
+        "fail_no_answer_meta": 0,
+        "fail_regex": 0,
+        "ok": 666
+      }
+    }
+  },
+  "thresholds": {
+    "angle_deg": 20.0,
+    "azimuth_deg": 20.0,
+    "distance_rel": 0.3,
+    "elevation_deg": 10.0,
+    "iou": 0.5,
+    "onset_s": 0.4
+  }
+}

runs/stage2_lower_lr/epoch_metrics.jsonl

@@ -0,0 +1 @@
+{"epoch": 1, "train_loss": 4.902002878892228, "train_supervised_tokens": 81176545.0, "optimizer_steps": 11339.0, "epoch_seconds": 39650.56145262718, "micro_steps": 22677.0, "valid_loss": 7.3201918202545215, "valid_supervised_tokens": 2296239.0, "valid_generate_examples": 32.0, "valid_exact_match": 0.0, "step_valid_subset_size": 1909.0, "valid_subset_size": 38186.0, "valid_generation_size": 32.0, "global_optimizer_step": 11339.0}

runs/stage2_lower_lr/processor/added_tokens.json

@@ -0,0 +1,25 @@
+{
+  "</tool_call>": 151658,
+  "<tool_call>": 151657,
+  "<|AUDIO|>": 151646,
+  "<|IMAGE|>": 151655,
+  "<|VIDEO|>": 151656,
+  "<|audio_bos|>": 151647,
+  "<|audio_eos|>": 151648,
+  "<|box_end|>": 151649,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|spatial|>": 151665,
+  "<|vision_bos|>": 151652,
+  "<|vision_eos|>": 151653,
+  "<|vision_pad|>": 151654
+}

runs/stage2_lower_lr/processor/chat_template.jinja

@@ -0,0 +1,7 @@
+{% set audio_count = namespace(value=0) %}{% set image_count = namespace(value=0) %}{% set video_count = namespace(value=0) %}{% for message in messages %}{% if loop.first and message['role'] != 'system' %}<|im_start|>system
+You are a helpful assistant.<|im_end|>
+{% endif %}<|im_start|>{{ message['role'] }}
+{% if message['content'] is string %}{{ message['content'] }}<|im_end|>
+{% else %}{% for content in message['content'] %}{% if content['type'] == 'image' or 'image' in content or 'image_url' in content %}{% set image_count.value = image_count.value + 1 %}{% if add_vision_id %}Picture {{ image_count.value }}: {% endif %}<|vision_bos|><|IMAGE|><|vision_eos|>{% elif content['type'] == 'audio' or 'audio' in content or 'audio_url' in content %}{% set audio_count.value = audio_count.value + 1 %}{% if add_audio_id %}Audio {{ audio_count.value }}: {% endif %}<|audio_bos|><|AUDIO|><|audio_eos|>{% elif content['type'] == 'video' or 'video' in content %}{% set video_count.value = video_count.value + 1 %}{% if add_vision_id %}Video {{ video_count.value }}: {% endif %}<|vision_bos|><|VIDEO|><|vision_eos|>{% elif 'text' in content %}{{ content['text'] }}{% endif %}{% endfor %}<|im_end|>
+{% endif %}{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant
+{% endif %}

runs/stage2_lower_lr/processor/preprocessor_config.json

@@ -0,0 +1,31 @@
+{
+  "chunk_length": 300,
+  "dither": 0.0,
+  "feature_extractor_type": "WhisperFeatureExtractor",
+  "feature_size": 128,
+  "hop_length": 160,
+  "image_mean": [
+    0.48145466,
+    0.4578275,
+    0.40821073
+  ],
+  "image_processor_type": "Qwen2VLImageProcessor",
+  "image_std": [
+    0.26862954,
+    0.26130258,
+    0.27577711
+  ],
+  "max_pixels": 12845056,
+  "merge_size": 2,
+  "min_pixels": 3136,
+  "n_fft": 400,
+  "n_samples": 4800000,
+  "nb_max_frames": 30000,
+  "padding_side": "right",
+  "padding_value": 0.0,
+  "patch_size": 14,
+  "processor_class": "Qwen2_5OmniSpatialProcessor",
+  "return_attention_mask": true,
+  "sampling_rate": 16000,
+  "temporal_patch_size": 2
+}

runs/stage2_lower_lr/processor/special_tokens_map.json

@@ -0,0 +1,32 @@
+{
+  "additional_special_tokens": [
+    {
+      "content": "<|spatial|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false
+    }
+  ],
+  "audio_bos_token": "<|audio_bos|>",
+  "audio_eos_token": "<|audio_eos|>",
+  "audio_token": "<|AUDIO|>",
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "image_token": "<|IMAGE|>",
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "video_token": "<|VIDEO|>",
+  "vision_bos_token": "<|vision_bos|>",
+  "vision_eos_token": "<|vision_eos|>"
+}

runs/stage2_lower_lr/processor/tokenizer_config.json

@@ -0,0 +1,218 @@
+{
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|AUDIO|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|audio_bos|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|audio_eos|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_bos|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_eos|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|IMAGE|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|VIDEO|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151665": {
+      "content": "<|spatial|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "additional_special_tokens": [
+    "<|spatial|>"
+  ],
+  "audio_bos_token": "<|audio_bos|>",
+  "audio_eos_token": "<|audio_eos|>",
+  "audio_token": "<|AUDIO|>",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {
+    "audio_bos_token": "<|audio_bos|>",
+    "audio_eos_token": "<|audio_eos|>",
+    "audio_token": "<|AUDIO|>",
+    "image_token": "<|IMAGE|>",
+    "video_token": "<|VIDEO|>",
+    "vision_bos_token": "<|vision_bos|>",
+    "vision_eos_token": "<|vision_eos|>"
+  },
+  "image_token": "<|IMAGE|>",
+  "model_max_length": 32768,
+  "pad_token": "<|endoftext|>",
+  "processor_class": "Qwen2_5OmniSpatialProcessor",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null,
+  "video_token": "<|VIDEO|>",
+  "vision_bos_token": "<|vision_bos|>",
+  "vision_eos_token": "<|vision_eos|>"
+}

runs/stage2_lower_lr/train_args.json

@@ -0,0 +1,68 @@
+{
+  "audio_feature_cache_manifest": "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen/qwen_audio_cache_ov1_v6c/manifest.json",
+  "audio_feature_cache_max_entries": 256,
+  "batch_size": 4,
+  "beats_checkpoint": "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats/checkpoints/spatial_beats_ov1_local_spatial_v2_exp/02_spatial/best.pt",
+  "beats_repo": "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/unilm/beats",
+  "device": "cuda:0",
+  "distributed": true,
+  "dtype": "bfloat16",
+  "epochs": 3,
+  "grad_accum_steps": 2,
+  "gradient_checkpointing": true,
+  "local_rank": 0,
+  "log_every": 10,
+  "lora_alpha": 32,
+  "lora_dropout": 0.05,
+  "lora_r": 16,
+  "lora_target_modules": [
+    "q_proj",
+    "k_proj",
+    "v_proj",
+    "o_proj",
+    "gate_proj",
+    "up_proj",
+    "down_proj"
+  ],
+  "lora_target_prefixes": [
+    "thinker.model"
+  ],
+  "lr": 1e-05,
+  "max_grad_norm": 1.0,
+  "max_train_samples": null,
+  "max_valid_samples": null,
+  "model_id": "/apdcephfs_cq10/share_1603164/user/schmittzhu/model/Qwen2.5-Omni-7B",
+  "num_workers": 0,
+  "optimizer_step_per_batch": false,
+  "output_dir": "./runs/stage2_lower_lr",
+  "persistent_workers": false,
+  "prefetch_factor": 2,
+  "projector_fp32": false,
+  "qa_root": "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline/prepared_datasets/ov1_qa_pairs_v6c",
+  "qa_roots": [
+    "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline/prepared_datasets/ov1_qa_pairs_v6c"
+  ],
+  "rank": 0,
+  "resume_checkpoint_path": "./runs/stage1/checkpoints/best_trainable.pt",
+  "resume_model_only": true,
+  "resume_tag": null,
+  "save_every_epoch": true,
+  "save_every_n_optimizer_steps": 1000,
+  "save_full_model": false,
+  "seed": 1234,
+  "spatial_qwen_repo": "/apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen",
+  "step_valid_subset_ratio": 0.05,
+  "train_mode": "encoder_lora",
+  "train_split": "train",
+  "valid_do_sample": false,
+  "valid_every_n_optimizer_steps": 1000,
+  "valid_generate_batch_size": 1,
+  "valid_generate_max_samples": 32,
+  "valid_max_new_tokens": 48,
+  "valid_num_beams": 1,
+  "valid_split": "valid",
+  "valid_subset_ratio": 0.1,
+  "warmup_ratio": 0.03,
+  "weight_decay": 0.01,
+  "world_size": 8
+}

runs/stage2_lower_lr/valid_predictions/epoch_001.jsonl

@@ -0,0 +1,32 @@
+{"epoch": 1, "pair_id": "valid/0000052", "task_name": "detect_time", "prompt": "Question: When is the thunderstorm audible in this clip?\n\nAnswer with only the time span.\nUse the format \"Xs to Ys\" with one decimal place.", "answer": "0.0s to 17.8s", "prediction": "00000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0000645", "task_name": "classify_source_class_choice", "prompt": "Question: Pick the correct event type for this clip.\nOptions:\nA. speech\nB. clock\nC. footsteps\nD. guitar\n\nAnswer with only the option letter.", "answer": "D", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0000750", "task_name": "detect_source_active_at_time", "prompt": "Question: Would an annotator mark the wind present at 5.3s?\n\nAnswer with only yes or no.", "answer": "yes", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0000932", "task_name": "classify_source_class_choice", "prompt": "Question: Which option matches the sound event in this recording?\nOptions:\nA. body sound\nB. gong\nC. dog\nD. footsteps\n\nAnswer with only the option letter.", "answer": "B", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0003540", "task_name": "detect_source_active_at_time", "prompt": "Question: Would an annotator mark the tool present at 1.6s?\n\nAnswer with only yes or no.", "answer": "no", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0005627", "task_name": "caption_source_inventory", "prompt": "Question: Determine the sound class in this recording.\n\nAnswer with semicolon-separated bracketed entries.\nUse the format \"[class_1]; [class_2]; ...\".\nKeep the class names short and lowercase.", "answer": "[glass]", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0006468", "task_name": "classify_elevation_bin_choice", "prompt": "Use the DCASE FOA coordinate system.\nThe +x axis points forward, +y points left, and +z points up.\nAzimuth increases counterclockwise when viewed from above.\nLeft corresponds to positive azimuth and right corresponds to negative azimuth.\nAzimuth is reported in degrees in [-180, 180], and elevation is reported in degrees in [-90, 90].\n\nQuestion: Which vertical angle bin matches the wind instrument?\nOptions:\nA. -50 to -30 degrees\nB. -30 to -10 degrees\nC. -10 to 10 degrees\nD. 70 to 90 degrees\n\nAnswer with only the option letter.", "answer": "B", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0010040", "task_name": "classify_distance_bin_choice", "prompt": "Interpret distance as the straight-line distance from the listener to the active sound source.\nDistances are grouped into coarse bins in meters.\n\nQuestion: Choose the distance category for this sound event.\nOptions:\nA. 2.8m to 3.9m meters\nB. 0.5m to 1.8m meters\nC. 3.9m to 6.5m meters\nD. 1.8m to 2.8m meters\n\nAnswer with only the option letter.", "answer": "C", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0010467", "task_name": "classify_azimuth_bin_choice", "prompt": "Use the DCASE FOA coordinate system.\nThe +x axis points forward, +y points left, and +z points up.\nAzimuth increases counterclockwise when viewed from above.\nLeft corresponds to positive azimuth and right corresponds to negative azimuth.\nAzimuth is reported in degrees in [-180, 180], and elevation is reported in degrees in [-90, 90].\n\nQuestion: Which azimuth interval should be assigned to the source?\nOptions:\nA. 160 to 180 degrees\nB. 40 to 60 degrees\nC. -160 to -140 degrees\nD. -120 to -100 degrees\n\nAnswer with only the option letter.", "answer": "B", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0011275", "task_name": "classify_azimuth_bin_choice", "prompt": "Use the DCASE FOA coordinate system.\nThe +x axis points forward, +y points left, and +z points up.\nAzimuth increases counterclockwise when viewed from above.\nLeft corresponds to positive azimuth and right corresponds to negative azimuth.\nAzimuth is reported in degrees in [-180, 180], and elevation is reported in degrees in [-90, 90].\n\nQuestion: Select the correct horizontal sector for the laughter.\nOptions:\nA. 100 to 120 degrees\nB. 140 to 160 degrees\nC. -80 to -60 degrees\nD. -160 to -140 degrees\n\nAnswer with only the option letter.", "answer": "A", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0011711", "task_name": "detect_source_active_at_time", "prompt": "Question: Is the musical instrument audible at 7.9s in this recording?\n\nAnswer with only yes or no.", "answer": "no", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0011785", "task_name": "classify_azimuth_bin_choice", "prompt": "Use the DCASE FOA coordinate system.\nThe +x axis points forward, +y points left, and +z points up.\nAzimuth increases counterclockwise when viewed from above.\nLeft corresponds to positive azimuth and right corresponds to negative azimuth.\nAzimuth is reported in degrees in [-180, 180], and elevation is reported in degrees in [-90, 90].\n\nQuestion: Select the azimuth range of the target sound source.\nOptions:\nA. -160 to -140 degrees\nB. 160 to 180 degrees\nC. -20 to 0 degrees\nD. 140 to 160 degrees\n\nAnswer with only the option letter.", "answer": "C", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0011903", "task_name": "detect_source_active_at_time", "prompt": "Question: Would an annotator mark the keyboard instrument present at 5.2s?\n\nAnswer with only yes or no.", "answer": "no", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0012503", "task_name": "classify_azimuth_bin_choice", "prompt": "Use the DCASE FOA coordinate system.\nThe +x axis points forward, +y points left, and +z points up.\nAzimuth increases counterclockwise when viewed from above.\nLeft corresponds to positive azimuth and right corresponds to negative azimuth.\nAzimuth is reported in degrees in [-180, 180], and elevation is reported in degrees in [-90, 90].\n\nQuestion: Which azimuth bin best matches the source location?\nOptions:\nA. -160 to -140 degrees\nB. -140 to -120 degrees\nC. -180 to -160 degrees\nD. 160 to 180 degrees\n\nAnswer with only the option letter.", "answer": "B", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0015163", "task_name": "classify_azimuth_bin_choice", "prompt": "Use the DCASE FOA coordinate system.\nThe +x axis points forward, +y points left, and +z points up.\nAzimuth increases counterclockwise when viewed from above.\nLeft corresponds to positive azimuth and right corresponds to negative azimuth.\nAzimuth is reported in degrees in [-180, 180], and elevation is reported in degrees in [-90, 90].\n\nQuestion: Which azimuth interval should be assigned to the source?\nOptions:\nA. 140 to 160 degrees\nB. 120 to 140 degrees\nC. -140 to -120 degrees\nD. -160 to -140 degrees\n\nAnswer with only the option letter.", "answer": "D", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0016118", "task_name": "detect_time", "prompt": "Question: What are the start and end timestamps of the target event?\n\nAnswer with only the time span.\nUse the format \"Xs to Ys\" with one decimal place.", "answer": "0.5s to 18.6s", "prediction": "00000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0018207", "task_name": "detect_source_active_at_time", "prompt": "Question: At 3.2s, is the target source active?\n\nAnswer with only yes or no.", "answer": "no", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0019842", "task_name": "detect_source_active_at_time", "prompt": "Question: Is the target class audible at 2.4s?\n\nAnswer with only yes or no.", "answer": "yes", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0019881", "task_name": "detect_source_active_at_time", "prompt": "Question: At 5.3s, should the machine be considered present?\n\nAnswer with only yes or no.", "answer": "yes", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0020552", "task_name": "classify_azimuth_bin_choice", "prompt": "Use the DCASE FOA coordinate system.\nThe +x axis points forward, +y points left, and +z points up.\nAzimuth increases counterclockwise when viewed from above.\nLeft corresponds to positive azimuth and right corresponds to negative azimuth.\nAzimuth is reported in degrees in [-180, 180], and elevation is reported in degrees in [-90, 90].\n\nQuestion: Choose the correct 20-degree azimuth bin for the guitar.\nOptions:\nA. 100 to 120 degrees\nB. -160 to -140 degrees\nC. -100 to -80 degrees\nD. -40 to -20 degrees\n\nAnswer with only the option letter.", "answer": "D", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0021505", "task_name": "classify_distance_bin_choice", "prompt": "Interpret distance as the straight-line distance from the listener to the active sound source.\nDistances are grouped into coarse bins in meters.\n\nQuestion: Which option marks the distance bin of the event?\nOptions:\nA. 3.9m to 6.5m meters\nB. 2.8m to 3.9m meters\nC. 0.5m to 1.8m meters\nD. 1.8m to 2.8m meters\n\nAnswer with only the option letter.", "answer": "B", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0021877", "task_name": "classify_azimuth_bin_choice", "prompt": "Use the DCASE FOA coordinate system.\nThe +x axis points forward, +y points left, and +z points up.\nAzimuth increases counterclockwise when viewed from above.\nLeft corresponds to positive azimuth and right corresponds to negative azimuth.\nAzimuth is reported in degrees in [-180, 180], and elevation is reported in degrees in [-90, 90].\n\nQuestion: Choose the source azimuth category.\nOptions:\nA. -20 to 0 degrees\nB. 140 to 160 degrees\nC. -80 to -60 degrees\nD. -120 to -100 degrees\n\nAnswer with only the option letter.", "answer": "D", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0024184", "task_name": "detect_source_active_at_time", "prompt": "Question: At 11.7s, is the target source active?\n\nAnswer with only yes or no.", "answer": "yes", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0025482", "task_name": "detect_time", "prompt": "Question: At what times does the zipper occur?\n\nAnswer with only the time span.\nUse the format \"Xs to Ys\" with one decimal place.", "answer": "3.0s to 16.2s", "prediction": "00000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0025719", "task_name": "detect_source_active_at_time", "prompt": "Question: Would you answer that the drum is active at 3.0s?\n\nAnswer with only yes or no.", "answer": "yes", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0029014", "task_name": "classify_elevation_bin_choice", "prompt": "Use the DCASE FOA coordinate system.\nThe +x axis points forward, +y points left, and +z points up.\nAzimuth increases counterclockwise when viewed from above.\nLeft corresponds to positive azimuth and right corresponds to negative azimuth.\nAzimuth is reported in degrees in [-180, 180], and elevation is reported in degrees in [-90, 90].\n\nQuestion: Select the source direction bin on the vertical axis.\nOptions:\nA. -90 to -70 degrees\nB. 10 to 30 degrees\nC. -50 to -30 degrees\nD. -10 to 10 degrees\n\nAnswer with only the option letter.", "answer": "C", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0029104", "task_name": "caption_source_inventory", "prompt": "Question: List the sound class present in this clip.\n\nAnswer with semicolon-separated bracketed entries.\nUse the format \"[class_1]; [class_2]; ...\".\nKeep the class names short and lowercase.", "answer": "[keyboard instrument]", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0029768", "task_name": "classify_distance_bin_choice", "prompt": "Interpret distance as the straight-line distance from the listener to the active sound source.\nDistances are grouped into coarse bins in meters.\n\nQuestion: Select the correct radial range for the sound source.\nOptions:\nA. 3.9m to 6.5m meters\nB. 2.8m to 3.9m meters\nC. 0.5m to 1.8m meters\nD. 1.8m to 2.8m meters\n\nAnswer with only the option letter.", "answer": "C", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0035090", "task_name": "classify_distance_bin_choice", "prompt": "Interpret distance as the straight-line distance from the listener to the active sound source.\nDistances are grouped into coarse bins in meters.\n\nQuestion: Choose the distance category for this sound event.\nOptions:\nA. 2.8m to 3.9m meters\nB. 3.9m to 6.5m meters\nC. 1.8m to 2.8m meters\nD. 0.5m to 1.8m meters\n\nAnswer with only the option letter.", "answer": "C", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0035987", "task_name": "detect_source_active_at_time", "prompt": "Question: Would you answer that the singing is active at 3.4s?\n\nAnswer with only yes or no.", "answer": "no", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0036316", "task_name": "classify_azimuth_bin_choice", "prompt": "Use the DCASE FOA coordinate system.\nThe +x axis points forward, +y points left, and +z points up.\nAzimuth increases counterclockwise when viewed from above.\nLeft corresponds to positive azimuth and right corresponds to negative azimuth.\nAzimuth is reported in degrees in [-180, 180], and elevation is reported in degrees in [-90, 90].\n\nQuestion: Pick the correct horizontal angle interval for the keyboard instrument.\nOptions:\nA. 20 to 40 degrees\nB. 40 to 60 degrees\nC. -20 to 0 degrees\nD. -180 to -160 degrees\n\nAnswer with only the option letter.", "answer": "C", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}
+{"epoch": 1, "pair_id": "valid/0037564", "task_name": "classify_source_class_choice", "prompt": "Question: Select the source class that best describes this segment.\nOptions:\nA. thunderstorm\nB. camera\nC. alarm\nD. tool\n\nAnswer with only the option letter.", "answer": "C", "prediction": "000000000000000000000000000000000000000000000000", "exact_match": 0}

runs/v13d_easy_llmqa_af3/stage1_projector/epoch_metrics.jsonl

@@ -0,0 +1,2 @@
+{"epoch": 1, "train_loss": 0.5010758844768967, "train_supervised_tokens": 16150791.0, "optimizer_steps": 16354.0, "epoch_seconds": 26248.89346075058, "micro_steps": 49060.0, "valid_loss": 0.5031168303269897, "valid_supervised_tokens": 1804356.0, "valid_generate_examples": 32.0, "valid_exact_match": 0.03125, "step_valid_subset_size": 4160.0, "valid_subset_size": 83193.0, "valid_generation_size": 32.0, "global_optimizer_step": 16354.0}
+{"epoch": 2, "train_loss": 0.43175358717016993, "train_supervised_tokens": 16150806.0, "optimizer_steps": 16354.0, "epoch_seconds": 26422.265491247177, "micro_steps": 49060.0, "valid_loss": 0.502653176573906, "valid_supervised_tokens": 1804356.0, "valid_generate_examples": 32.0, "valid_exact_match": 0.0625, "step_valid_subset_size": 4160.0, "valid_subset_size": 83193.0, "valid_generation_size": 32.0, "global_optimizer_step": 32708.0}

runs/v13d_easy_llmqa_af3/stage1_projector/processor/chat_template.jinja

@@ -0,0 +1,6 @@
+{% if messages[0]['role'] != 'system' %}<|im_start|>system
+You are a helpful assistant.<|im_end|>
+{% endif %}{% for m in messages if m['content'] is not none %}<|im_start|>{{ m['role'] }}
+{% if m['content'] is string %}{{ m['content'] }}{% else %}{% set audio = namespace(found=False) %}{% set text_buf = namespace(v='') %}{% for c in m['content'] %}{% if c.get('type') == 'audio' or 'audio' in c %}{% set audio.found = True %}{% elif c.get('type') == 'text' or 'text' in c %}{% set text_buf.v = text_buf.v + c['text'] %}{% endif %}{% endfor %}{% if audio.found %}{{ '<sound>' }}{% endif %}{{ text_buf.v }}{% endif %}<|im_end|>
+{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant
+{% endif %}

runs/v13d_easy_llmqa_af3/stage1_projector/processor/processor_config.json

@@ -0,0 +1,23 @@
+{
+  "audio_token": "<sound>",
+  "default_transcription_prompt": "Transcribe the input speech.",
+  "feature_extractor": {
+    "chunk_length": 30,
+    "dither": 0.0,
+    "feature_extractor_type": "WhisperFeatureExtractor",
+    "feature_size": 128,
+    "hop_length": 160,
+    "n_fft": 400,
+    "n_samples": 480000,
+    "nb_max_frames": 3000,
+    "padding_side": "right",
+    "padding_value": 0.0,
+    "return_attention_mask": true,
+    "sampling_rate": 16000
+  },
+  "max_audio_len": 600,
+  "processor_class": "AudioFlamingo3SpatialProcessor",
+  "spatial_audio_max_seconds": 20.0,
+  "spatial_beats_target_token_rate": 2.5,
+  "spatial_token": "<spatial>"
+}

runs/v13d_easy_llmqa_af3/stage1_projector/processor/tokenizer_config.json

@@ -0,0 +1,20 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": "[BOS]",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": [
+    "<spatial>"
+  ],
+  "is_local": true,
+  "legacy": false,
+  "model_max_length": 32768,
+  "pad_token": "[PAD]",
+  "padding_side": "right",
+  "processor_class": "AudioFlamingo3SpatialProcessor",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}