code/README.md

# EchoScout four-task search code

Portable source release for the EchoScout four-task benchmark. This archive contains implementation,
configuration, dependency specifications, and synthetic code checks only.
The data archives in the parent repository provide the matching dataset. Model
weights, deployment credentials, logs, and repository history are not bundled.

## Contents

- `omini/`: binaural feature extraction, temporal audio encoder (41 tokens),
  normalization helpers, and canonical waveform preprocessing.
- `training/stereo_backbones_v1/`: Qwen3.5-9B and InternVL3.5-8B extension,
  audio alignment, standalone SFT, and audiovisual inverse dynamics (AV-IDM).
- `training/hvs_visual_to_audio_v1/`: HVS-3B extension from visual SFT+IDM
  initialization, including the validation retention gate.
- `training/vision_sft_idm_v1/`: HVS visual initialization training.
- `training/echoscout_no_sft_v1/`: matched ablation removing only standalone SFT.
- Other training subdirectories: shared data construction and objective utilities.
- `evaluation/fourtask_common_v3/`: semantic prompts, legal action parser,
  local model registry, and input protocol.
- `tools/`: panorama projection, deterministic scoring, and file utilities.

Directory version suffixes are import-compatible implementation namespaces;
they do not contain historical run artifacts.

## Environment

Use Linux, Python 3.11 or later, and a CUDA GPU supporting BF16. Install a
CUDA-compatible PyTorch build and then the dependencies:

```bash
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
export PYTHONPATH="$PWD:$PWD/tools"
python tests/check_code.py
python tests/check_numeric.py
```

Pinned versions reflect the source training runtime. Only CPU code checks,
not full GPU training or checkpoint inference, were run on this portable
export. Full-parameter training uses host-side FP32 optimizer state; the
configuration retains the original memory and storage requirements. No
automatic GPU allocation, remote access, downloading, or job scheduler is
included. Run commands from the extracted package root.

## External inputs

Supply local model directories in `configs/models.json`. They must contain
the corresponding native pretrained model and processor files. Loading uses
`local_files_only=True`; no credentials or automatic model downloads are used.
The HVS entry expects the HVS visual backbone, not an arbitrary Qwen checkpoint.

Supply the pretrained binaural encoder as
`checkpoints/audio_encoder/encoder_best.pt`. Its architecture metadata is in
`configs/audio_encoder.json` (355 semantic classes, two temporal layers).
This is an initialization dependency and is intentionally not bundled.
The checkpoint is loaded strictly; a differently trained or randomly
initialized encoder is not an equivalent replacement.

Extract the matching data archives from the parent dataset repository with
`python ../scripts/extract_data.py --destination .`.
`DATA_FORMAT.md` describes expected locations and fields. Dataset size constants
are protocol specifications, not experimental outcomes.

For the HVS branch, supply a verified visual SFT+IDM initialization checkpoint,
or produce one using the included initialization code:

```bash
python training/vision_sft_idm_v1/train_sft.py --model hvs_original --gpu 0
python training/vision_sft_idm_v1/train_idm.py --model hvs_original --gpu 0
python configure.py --bind-visual
```

`--bind-visual` verifies the initializer against its `complete.json` and binds
its SHA256 in the HVS configuration. It does not select a checkpoint from test
results. Without this step the placeholder hash deliberately prevents using
an unverified initializer. The no-SFT HVS ablation retains this visual ancestor.

## Data preparation

With source assets and annotations present:

```bash
python prepare_data.py merge
python prepare_data.py policy
python prepare_data.py auxiliary
```

These create policy rows, visibility-gated expert trajectories, paired temporal
counterfactuals, and integrity manifests under `workspace/`. The builders reject
train/validation/test scene or clip overlap. They do not establish disjointness
of all underlying assets or pretrained-model exposure.

## Training

Edit portable paths in the relevant `config.json` files before starting a run.
After configuration changes, refresh the matched-ablation integrity hashes:

```bash
python configure.py
python training/echoscout_no_sft_v1/train.py --model qwen35_9b --verify-only
```

The default full method uses connector-only alignment followed by
full-parameter SFT and AV-IDM, without LoRA. AV-IDM retains the joint search
policy loss. Training budgets, learning rates, validation rules, task sampling,
and loss weights are specified in the source/configuration.

```bash
python training/stereo_backbones_v1/train.py --model qwen35_9b --gpu 0 --smoke-only
python training/stereo_backbones_v1/train.py --model qwen35_9b --gpu 0
python training/stereo_backbones_v1/train.py --model internvl35_8b --gpu 0
python training/hvs_visual_to_audio_v1/train.py --model hvs_original --gpu 0
```

The smoke option performs real per-task backward checks after inputs are
installed. It is not a substitute for the full training budget. Partial runs
are deliberately not silently restarted with a fresh optimizer.

Matched no-standalone-SFT ablation:

```bash
python training/echoscout_no_sft_v1/train.py --model qwen35_9b --gpu 0
python training/echoscout_no_sft_v1/train.py --model internvl35_8b --gpu 0
python training/echoscout_no_sft_v1/train.py --model hvs_original --gpu 0
```

This wrapper verifies local source hashes, removes only the standalone SFT
stage, and retains alignment, initialization, and AV-IDM policy supervision.
Hashes are for this portable source package, not the original deployment.

## Evaluation

Training completion and a nonempty final checkpoint are required. For example:

```bash
python training/stereo_backbones_v1/evaluate.py --model stereo_qwen35_9b_idm --suite static --gpu 0
python training/stereo_backbones_v1/evaluate.py --model stereo_qwen35_9b_idm --suite sound --gpu 0
python training/hvs_visual_to_audio_v1/evaluate.py --model visual_to_audio_hvs_original_idm --suite sound --gpu 0
python training/echoscout_no_sft_v1/evaluate.py --model echoscout_no_sft_internvl35_8b --suite sound --gpu 0
```

Run both suites for each model. `static` covers HOS; `sound` covers category,
duration, and temporal-order search. Inference outputs would be written under
`runs/` when executed; this directory is not included in the archive.
The runner verifies exact episode IDs and independently replays actions and
recomputes scores on completion. Intermediate counts are not final results.

Qwen/InternVL input-removal evaluation uses:

```bash
python training/stereo_backbones_v1/ablation.py --model stereo_qwen35_9b_idm --ablation no_audio --suite sound --gpu 0
```

`no_image` is also supported. Every invocation performs its own inference;
no static results are bundled or reused.

## Protocol and limitations

The search policy sees the current local view plus at most two historical
views/actions, a text instruction, and (for sound tasks) the complete fixed
10-second stereo recording. Camera rotation does not rotate the microphone
recording. There are at most ten decisions including submission; invalid
outputs consume a decision without changing pose. Generation is greedy with
512 new tokens. Native backbone image/token serialization can differ even
when source observations and semantic prompts are identical.

Canonical per-channel RMS normalization preserves channel ordering and phase
structure but removes absolute inter-channel level differences. Auxiliary
counterfactual recordings are dry mixtures duplicated across channels; they
teach temporal semantics and do not simulate binaural spatial acoustics.
Visibility and sound scoring use annotation-derived geometric proxies.
Checkpoint selection uses validation, never test performance.

## Export changes

Deployment paths were replaced with package-relative configuration. The
registry was reduced to the supported backbones, unrelated schedulers and
recovery exceptions were omitted, shared utilities were extracted, and
matched-ablation hashes were regenerated. Numerical preprocessing, model
modules, objectives, action parsing, and scoring were retained. Public
third-party attribution remains in `THIRD_PARTY_NOTICES.md` and `licenses/`.