# Data Format

All files use JSON Lines: each line is one independent JSON object.

## Files and Statistics

| File | Examples | Mean input tokens | Median input tokens | SHA256 |
|---|---:|---:|---:|---|
| `H2S-SFT.jsonl` | 4,228 | 44,097 | 28,304 | `c0a97173d03eb08585b384e8bb053d7b9f428876a6089538e3b458106f796359` |
| `H2S-RL.jsonl` | 2,419 | 43,656 | 27,726 | `5deadcab27337660cf79f3d481ed11385529e54edf1671f1088711a63ee956c5` |
| `H2S-Bench.jsonl` | 2,575 | 44,047 | 24,691 | `ff5daf469ea7ff71ec7a1ebf89bfa33a06a844546b714546f0fbe7d576b3980b` |

Common fields are:

| Field | Type | Description |
|---|---|---|
| `schema_version` | string | Record schema identifier |
| `id` | string | Stable example identifier |
| `benchmark` | string | Source benchmark/task family |
| `ability` | string | Long-context capability category |
| `question` | string | Question or task instruction |
| `input_token_num` | integer | Input tokens after preprocessing |
| `input_token_num_original` | integer | Input tokens before preprocessing |

## H2S-SFT

Schema version: `v3_sft_v1`.

Every record contains a three-turn `messages` list with `system`, `user`, and
`assistant` roles. The assistant target follows:

```text
<evidence>source-addressable evidence</evidence>
<summary>question-conditioned summary</summary>
<answer>benchmark-native final answer</answer>
```

```json
{
  "schema_version": "v3_sft_v1",
  "id": "...",
  "benchmark": "...",
  "ability": "...",
  "question": "...",
  "input_token_num": 32768,
  "input_token_num_original": 32000,
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "<evidence>...</evidence>..."}
  ]
}
```

## H2S-RL

Schema version: `v3_rl_v1`.

Each record contains a two-turn `prompt`, benchmark-native `gt`, and process
supervision used by the deterministic H2S reward:

| Field | Type | Description |
|---|---|---|
| `prompt` | list | System and user messages supplied to the policy |
| `gt` | object | Final-answer target and source metadata |
| `reference_evidence` | object | Claims mapped to source block/span locators |
| `reference_spans` | list | Block IDs, document IDs, offsets, and exact spans |
| `reference_claims` | list | Atomic claims and their supporting evidence |
| `reference_subqueries` | list | Decomposed information needs and linked claims |
| `reference_summary` | string | Question-conditioned reference summary |
| `reward_metadata` | object | Answerability, renderer, and quality-control metadata |

## H2S-Bench

Schema version: `v3_eval_v1`.

```json
{
  "schema_version": "v3_eval_v1",
  "id": "...",
  "benchmark": "MRCR",
  "ability": "precise_retrieval",
  "question": "...",
  "gt": {"answer": "..."},
  "input_token_num": 32768,
  "input_token_num_original": 32000,
  "input_token_num_mode": "qwen2.5-7b-instruct-chat-template",
  "input": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."}
  ]
}
```

`gt.answer` preserves each benchmark's native representation and can be a
string or a list. H2S-Bench contains final-answer references only; it does not
contain reference evidence, spans, claims, subqueries, or summaries.