| # Parason: Subtask- and Trial-Parallel Reasoning Data |
|
|
| Training data for **Parason**, which turns long sequential reasoning traces into *executable parallel |
| programs*. Three subsets, one per training stage. |
|
|
| | Subfolder | Stage | Contents | |
| | :--- | :--- | :--- | |
| | `sft-8x` | SFT (main) | Parallel trajectories with `<Subtask>` / `<Trial>` branches, Trial stages expanded 8x — extra branches sampled with Qwen3-8B, their `<Trial>` goals written by Gemini-3-Flash | |
| | `sft-1x` | SFT (baseline) | The unexpanded parallel trajectories: 964 annotated Qwen3-8B traces released by ThreadWeaver, each parallel stage labeled Subtask or Trial by Gemini-3-Flash | |
| | `rl` | PA-GRPO | RL prompt set — [Polaris-53k](https://github.com/ChenxinAn-fdu/POLARIS) (~53,000 complex reasoning problems), used unmodified | |
|
|
| Each subfolder holds a single `train.parquet`; there is no validation split. The two SFT sets are |
| matched in effective token budget (8 epochs over `sft-1x` ≈ 1 epoch over `sft-8x`). |
|
|
| ## Trajectory format |
|
|
| The SFT trajectories mark two distinct forms of parallelism: |
|
|
| * **`<Subtask>`** — AND-branches. The problem splits into independent sub-goals, each producing a |
| *required* result. Shortens the critical path. |
| * **`<Trial>`** — OR-branches. Competing speculative attempts test different hypotheses; all outputs |
| are concatenated back into the reasoning history. Widens search. |
|
|
| ``` |
| <think> |
| [Initial problem analysis and setup] |
| |
| <Parallel> |
| <Outlines> |
| <Trial> 1: [First hypothesis to test] |
| <Trial> 2: [Second hypothesis to test] |
| <Subtask> 3: [An independent sub-goal whose result is required] |
| </Outlines> |
| |
| <Thread> |
| 1: [Detailed reasoning for the first branch...] |
| </Thread> |
| <Thread> |
| 2: [Detailed reasoning for the second branch...] |
| </Thread> |
| <Thread> |
| 3: [Detailed reasoning for the third branch...] |
| </Thread> |
| </Parallel> |
| </think> |
| |
| [Final answer in \boxed{} format] |
| ``` |
|
|
| Reserved tags, added as special tokens before training: |
|
|
| ``` |
| <Think> </Think> <Parallel> </Parallel> <Outlines> </Outlines> <Outline> </Outline> |
| <Trial> </Trial> <Subtask> </Subtask> <Thread> </Thread> <Conclusion> </Conclusion> |
| ``` |
|
|
| ## Schema |
|
|
| **`sft-8x`, `sft-1x`** |
|
|
| | Field | Type | Description | |
| | :--- | :--- | :--- | |
| | `uuid` | `string` | Source-trace identifier | |
| | `question` | `string` | Problem statement | |
| | `response_reasoning` | `string` | Trajectory body — the `<think>` contents, containing the `<Parallel>` blocks | |
| | `response_content` | `string` | Post-`</think>` final answer, ending in `\boxed{...}` | |
| | `correctness` | `bool` | Whether the source trace was graded correct | |
| | `raw_messages` | `list[dict]` | `[{"role": "user", ...}, {"role": "assistant", ...}]` | |
| | `qwen_text` | `string` | **The training field.** Full trace with the Qwen3 chat template applied | |
| | `num_qwen_tokens` | `int` | Token count of `qwen_text` under the Qwen3-8B tokenizer | |
|
|
| **`rl`** — standard [veRL](https://github.com/volcengine/verl) prompt format: `prompt` |
| (`list[dict]`), `data_source`, `ability`, `reward_model` (`{"style": "rule", "ground_truth": ...}`), |
| `extra_info`. |
|
|
|
|
