# RATs40K Adapter for ChatTime

This folder adapts ChatTime to the RATs40K univariate anomaly QA task.

It intentionally uses the numeric `Observation` field only. It does not use
`FigurePath`, so the resulting baseline should be compared with numeric-only
Time-RA settings rather than VLM image-input settings.

## Required inputs

- `MODEL_PATH`: local ChatTime model directory. This is required by default.
- `PYTHON_BIN`: Python executable. The shell scripts default to
  `/dev/shm/suiqk/conda_envs/scalerag-ts-v4/bin/python`.
- `DATA_PATH`: defaults to
  `/mnt/share01/sqk/datasets/RATs40K/RATs-Uni-TSImage_Reason.json`.

The scripts do not download HuggingFace weights unless `ALLOW_HF_DOWNLOAD=1`
is set explicitly.

The default precision is FP16 because the configured four-GPU machine uses
Tesla V100 GPUs. SFT defaults to regular FP16 LoRA with
`LOAD_IN_4BIT=0`, `PER_DEVICE_TRAIN_BATCH_SIZE=1`, and
`GRADIENT_ACCUMULATION_STEPS=16`.

Evaluation defaults to `EVAL_BATCH_SIZE=4` per GPU. With four GPUs, the
maximum global evaluation batch size is 16.

The task prompt is aligned with Time-RA's univariate
`USER_DETECTION_PROMPT`; ChatTime still receives the normalized/discretized
series through its native `### Input` section. Evaluation checks prompt token
lengths and fails instead of truncating. Defaults are `MAX_INPUT_TOKENS=3936`,
`MAX_NEW_TOKENS=160`, and `MAX_SEQ_LENGTH=4096` for SFT.

## Zero-shot

```bash
cd /mnt/share01/sqk/ChatTime
MODEL_PATH=/mnt/share01/sqk/models/ChatTime-1-7B-Chat \
bash rats40k_adapter/run_zeroshot_4gpu.sh
```

## SFT + Eval

```bash
cd /mnt/share01/sqk/ChatTime
MODEL_PATH=/mnt/share01/sqk/models/ChatTime-1-7B-Chat \
bash rats40k_adapter/run_sft_4gpu.sh
```

## Zero-shot Then SFT + Eval

```bash
cd /mnt/share01/sqk/ChatTime
bash rats40k_adapter/run_zeroshot_then_sft_4gpu.sh
```

## Saved Results

- Zero-shot outputs:
  `/mnt/share01/sqk/ChatTime/rats40k_adapter/outputs/pipeline_20260608_175250/zeroshot`
- SFT outputs:
  `/mnt/share01/sqk/ChatTime/rats40k_adapter/outputs/pipeline_20260608_175250/sft`

Useful smoke-test knobs:

```bash
MAX_TRAIN_SAMPLES=128 MAX_EVAL_SAMPLES=64 bash rats40k_adapter/run_zeroshot_then_sft_4gpu.sh
MAX_TRAIN_SAMPLES=128 MAX_EVAL_SAMPLES=64 bash rats40k_adapter/run_sft_4gpu.sh
MAX_EVAL_SAMPLES=64 bash rats40k_adapter/run_zeroshot_4gpu.sh
```