README.md

---
configs:
- config_name: default
  data_files:
  - split: train_no_aux
    path: data/train_no_aux-*
  - split: train_aux_cascade
    path: data/train_aux_cascade-*
  - split: train_aux_multitask
    path: data/train_aux_multitask-*
  - split: test
    path: data/test-*
  - split: validation
    path: data/validation-*
dataset_info:
  features:
  - name: data_source
    dtype: string
  - name: prompt
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  - name: ability
    dtype: string
  - name: reward_model
    struct:
    - name: style
      dtype: string
    - name: extraction_method
      dtype: string
    - name: ground_truth
      dtype: large_string
    - name: key
      dtype: string
  - name: extra_info
    struct:
    - name: id
      dtype: string
    - name: lower_pass_rate
      dtype: float64
    - name: upper_pass_rate
      dtype: float64
  splits:
  - name: train_no_aux
    num_bytes: 11789370808
    num_examples: 9693
  - name: train_aux_cascade
    num_bytes: 11846414523
    num_examples: 25538
  - name: train_aux_multitask
    num_bytes: 11846423932
    num_examples: 25538
  - name: test
    num_bytes: 289630396
    num_examples: 175
  - name: validation
    num_bytes: 985682185
    num_examples: 481
  download_size: 36758617513
  dataset_size: 36757521844
---

# FinalMix2

A multi-task **code reinforcement-learning** dataset mixture in the
[`verl`](https://github.com/volcengine/verl) RL prompt format. It pairs a
code-generation split with a suite of auxiliary code-understanding tasks so the
same corpus can drive three training regimes from one repo. It is the
V3-dedupe successor to `OctoReasoner/FinalMix` (see
[Relationship to FinalMix (v1)](#relationship-to-finalmix-v1)).

## Splits

| Split | Rows | Contents | Use |
|-------|-----:|----------|-----|
| `train_no_aux` | 9,693 | code-generation only | RL on code gen alone |
| `train_aux_cascade` | 25,538 | all 15,845 auxiliary rows first, then the 9,693 code rows appended (order preserved) | cascade / curriculum RL (aux → code) |
| `train_aux_multitask` | 25,538 | the same code + aux rows concatenated and shuffled (`seed=42`) | mixed multi-task RL |
| `validation` | 481 | held-out code-generation problems | eval |
| `test` | 175 | LiveCodeBench-v6 problems | eval |

The three training splits are built from the **same** underlying rows — they
differ only in which tasks are included and in what order — so they form a
controlled three-way comparison:

1. **`train_no_aux`** — code generation only.
2. **`train_aux_cascade`** — auxiliary tasks then code, for cascade RL.
3. **`train_aux_multitask`** — code and auxiliary tasks interleaved, for mixed multi-task RL.

```python
from datasets import load_dataset

code_only  = load_dataset("OctoReasoner/FinalMix2", split="train_no_aux")
cascade    = load_dataset("OctoReasoner/FinalMix2", split="train_aux_cascade")
multitask  = load_dataset("OctoReasoner/FinalMix2", split="train_aux_multitask")
val        = load_dataset("OctoReasoner/FinalMix2", split="validation")
test       = load_dataset("OctoReasoner/FinalMix2", split="test")
```

## Code split (9,693)

A more liberal ("V3") deduplication of the source code pools, rebalanced away
from the contest-heavy v1 mix toward PrimeIntellect:

| Source | Rows | Share |
|--------|-----:|------:|
| `code_primeintellect` | 5,241 | 54.1% |
| `code_contests_o` | 2,538 | 26.2% |
| `code_taco` | 1,721 | 17.8% |
| `code_lcbv5` | 193 | 2.0% |

## Auxiliary tasks (15,845)

Twelve `data_source`s spanning ~24 ability sub-tasks that probe code
understanding beyond generation:

- **Input/output reasoning** — `code_io_taco`, `code_functional_identity`
  (predict outputs from inputs / inputs from outputs, direct and MCQ).
- **Complexity** — `code_time_complexity`, `code_space_complexity`,
  `code_cpu_ranking`, `code_memory_ranking` (predict/rank time, space, CPU, memory).
- **Security** — `code_sast_cwe` (predict/localize CWE weaknesses).
- **Retrieval** — `code_crp_retrieval` (`coderpile_retrieval`).
- **Localization** — `code_change_localization`, `code_var_tracing`
  (locate edits; trace variable values).
- **Compilation** — `code_compile_status` (predict whether code compiles).
- **Instruction following** — `codeif` (verifiable instruction-following, generate & edit).

## Schema

Standard `verl` RL fields:

| Field | Type | Notes |
|-------|------|-------|
| `data_source` | string | routes the reward function |
| `prompt` | list of `{role, content}` | chat-formatted problem |
| `ability` | string | task category |
| `reward_model` | struct `{style, extraction_method, ground_truth, key}` | scoring spec |
| `extra_info` | struct `{id, lower_pass_rate, upper_pass_rate}` | per-example metadata |

Code-generation rows are scored by executing model output against tests in a
sandbox; auxiliary rows are scored by rule / answer extraction against
`ground_truth`.

## Relationship to FinalMix (v1)

`FinalMix2` rebuilds the code split of `OctoReasoner/FinalMix` on a more liberal
dedupe (9,693 code rows vs. 6,000) and rebalances the source distribution — v1
was `code_contests_o`-dominated (~50%), v2 leads with `code_primeintellect`
(~54%). The combined training splits grow accordingly (25,538 vs. 22,000). The
schema, the auxiliary-task set, and the `validation`/`test` eval splits are
carried over unchanged from v1.