---
configs:
  - config_name: facts
    data_files:
      - split: train
        path: facts/train_all_facts.jsonl
      - split: validation
        path: facts/validation_all_facts.jsonl
      - split: test1
        path: facts/test1_all_facts.jsonl
      - split: test2
        path: facts/test2_all_facts.jsonl
  - config_name: kto
    data_files:
      - split: train
        path: kto/train_kto_chat_format.jsonl
      - split: validation
        path: kto/validation_kto_chat_format.jsonl
      - split: test1
        path: kto/test1_kto_chat_format.jsonl
      - split: test2
        path: kto/test2_kto_chat_format.jsonl
  - config_name: dpo
    data_files:
      - split: train
        path: dpo/train_dpo_chat_format.jsonl
      - split: validation
        path: dpo/validation_dpo_chat_format.jsonl
      - split: test1
        path: dpo/test1_dpo_chat_format.jsonl
      - split: test2
        path: dpo/test2_dpo_chat_format.jsonl
  - config_name: sft2
    data_files:
      - split: train
        path: viewer/sft2/train_sft2_chat_format.jsonl
      - split: validation
        path: viewer/sft2/validation_sft2_chat_format.jsonl
      - split: test1
        path: viewer/sft2/test1_sft2_chat_format.jsonl
      - split: test2
        path: viewer/sft2/test2_sft2_chat_format.jsonl
  - config_name: kto_base
    data_files:
      - split: train
        path: kto_base/train_kto_chat_format.jsonl
      - split: validation
        path: kto_base/validation_kto_chat_format.jsonl
      - split: test1
        path: kto_base/test1_kto_chat_format.jsonl
      - split: test2
        path: kto_base/test2_kto_chat_format.jsonl
  - config_name: dpo_base
    data_files:
      - split: train
        path: dpo_base/train_dpo_chat_format.jsonl
      - split: validation
        path: dpo_base/validation_dpo_chat_format.jsonl
      - split: test1
        path: dpo_base/test1_dpo_chat_format.jsonl
      - split: test2
        path: dpo_base/test2_dpo_chat_format.jsonl
  - config_name: kto_sft1
    data_files:
      - split: train
        path: kto_sft1/train_kto_chat_format.jsonl
      - split: validation
        path: kto_sft1/validation_kto_chat_format.jsonl
      - split: test1
        path: kto_sft1/test1_kto_chat_format.jsonl
      - split: test2
        path: kto_sft1/test2_kto_chat_format.jsonl
  - config_name: dpo_sft1
    data_files:
      - split: train
        path: dpo_sft1/train_dpo_chat_format.jsonl
      - split: validation
        path: dpo_sft1/validation_dpo_chat_format.jsonl
      - split: test1
        path: dpo_sft1/test1_dpo_chat_format.jsonl
      - split: test2
        path: dpo_sft1/test2_dpo_chat_format.jsonl
  - config_name: kto_sft2
    data_files:
      - split: train
        path: kto_sft2/train_kto_chat_format.jsonl
      - split: validation
        path: kto_sft2/validation_kto_chat_format.jsonl
      - split: test1
        path: kto_sft2/test1_kto_chat_format.jsonl
      - split: test2
        path: kto_sft2/test2_kto_chat_format.jsonl
  - config_name: dpo_sft2
    data_files:
      - split: train
        path: dpo_sft2/train_dpo_chat_format.jsonl
      - split: validation
        path: dpo_sft2/validation_dpo_chat_format.jsonl
      - split: test1
        path: dpo_sft2/test1_dpo_chat_format.jsonl
      - split: test2
        path: dpo_sft2/test2_dpo_chat_format.jsonl
---

# MTS-Dialog RL Training Data

KTO/DPO/fact-extraction data for patient simulation on MTS-Dialog.
Generated with Kimi K2.6 (facts) and Qwen 3.5-4B / miquel_final_sft (negatives).

| Config | Content | Rows (train / val / test1 / test2) |
|--------|---------|------------------------------------|
| `facts` | 3-seed patient fact extractions | 1200 / 100 / 200 / 200 |
| `kto` | KTO pairs (gold + Qwen3.5-4B negative) | 10658 / 764 / 1602 / 1824 |
| `kto_base` | KTO pairs (gold + base Qwen negative, pre-remap fix) | 10658 / 764 / 1602 / 1824 |
| `kto_sft1` | KTO pairs (gold + SFT1 LoRa negative) | same |
| `kto_sft2` | KTO pairs (gold + miquel_final_sft negative) | same |
| `dpo` | DPO chosen/rejected (Qwen3.5-4B rejected) | 5329 / 382 / 801 / 912 |
| `dpo_base` | DPO chosen/rejected (base Qwen rejected, pre-remap fix) | 5329 / 382 / 801 / 912 |
| `dpo_sft1` | DPO chosen/rejected (SFT1 LoRa rejected) | same |
| `dpo_sft2` | DPO chosen/rejected (miquel_final_sft rejected) | same |
| `sft2` | Gold-only SFT data (new prompt + facts) | 5329 / 382 / 801 / 912 |