corbanyax's picture
Initial release: adapters/data, card and cross-references
dfb3de0 verified
|
Raw History Blame Contribute Delete
4.97 kB
---
license: cc0-1.0
language:
- en
pretty_name: Countdown-3 Puzzles with Hash-Defined Splits
size_categories:
- 100K<n<1M
task_categories:
- text-generation
- reinforcement-learning
tags:
- countdown
- arithmetic
- math-reasoning
- reinforcement-learning
- grpo
- verifiable-rewards
- synthetic
configs:
- config_name: default
data_files:
- split: pool
path: data/pool.parquet
- split: dev
path: data/dev.parquet
- split: dev200
path: data/dev200.parquet
- split: test_id
path: data/test_id.parquet
- split: test_ood_target
path: data/test_ood_target.parquet
- split: test_ood_4
path: data/test_ood_4.parquet
---
# Countdown-3 Puzzles with Hash-Defined Splits
**Countdown-3** arithmetic puzzles for RL post-training with verifiable rewards: given three integers in [1, 50]
and a target, write an expression that uses every number exactly once with `+ - * /` and parentheses and equals the
target. The universe is **enumerated exhaustively** with exact rational arithmetic (no sampling), and splits are
defined by **SHA-256 of the number multiset**, so held-out splits use number combinations never seen in training.
Used to train and evaluate [Dress Rehearsal](https://huggingface.co/corbanyax/Dress-Rehearsal-Qwen3-0.6B-4bit-KV-LoRA) and the
[KV-cache precision study](https://huggingface.co/corbanyax/KV-Cache-Precision-GRPO-Qwen3-0.6B-LoRA).
## Splits
| Split | Rows | Content |
|---|---|---|
| `pool` | 102,497 | training pool: every reachable integer target in [10, 100] for the remaining multisets |
| `dev` | 500 | first 500 multisets in hash order, one instance each |
| `dev200` | 200 | the first 200 problems of `dev`, the evaluation set used throughout the paper |
| `test_id` | 1,000 | next 1,000 multisets; evaluated **once**, for final policies only |
| `test_ood_target` | 1,000 | next 1,000 multisets with targets in [101, 200] |
| `test_ood_4` | 1,000 | 4 numbers in [1, 25], targets in [10, 100] |
Columns: `prompt_id` (e.g. `cd3:1-1-31=29`), `numbers` (list of int), `target` (int), `n_numbers` (int), `split`.
```python
from datasets import load_dataset
ds = load_dataset("corbanyax/Countdown-3-Hash-Splits") # all splits
dev200 = load_dataset("corbanyax/Countdown-3-Hash-Splits", split="dev200")
```
## Construction
3-number multisets (22,100) are ordered by SHA-256 of `cd3:a-b-c`. The first 500 form `dev`, the next 1,000
`test_id`, the next 1,000 `test_ood_target`; each contributes exactly one instance, the reachable target whose
instance key has the smallest SHA-256. All remaining multisets form `pool`. 4-number multisets are ordered the same
way and the first 1,000 form `test_ood_4`. Reachable values come from exact rational arithmetic over every split into
two sub-multisets; unary minus is not allowed. Rebuilding gives byte-identical files (SHA-256 of `countdown_3.parquet`
recorded in the project log).
## Prompt and reward
System: `You are a careful assistant that solves arithmetic puzzles.` User: `Using the numbers {numbers}, create an arithmetic expression that equals
{target}. Use every number exactly once. You may use +, -, *, / and parentheses. Show brief reasoning. End with the
expression alone between answer tags, for example <answer>(100 - 75) * 60</answer>`. The format example uses numbers
above 50, so it never matches an instance. Reward 1 if the `<answer>` expression uses each number exactly once and
evaluates exactly to the target, else 0; no format credit.
The original per-file layout is kept under `raw/` (`countdown_3.parquet`, `countdown_4.parquet`).
## Project family
| Item | Hugging Face | Kaggle mirror |
|---|---|---|
| Dress Rehearsal adapters (this method) | [corbanyax/Dress-Rehearsal-Qwen3-0.6B-4bit-KV-LoRA](https://huggingface.co/corbanyax/Dress-Rehearsal-Qwen3-0.6B-4bit-KV-LoRA) | [model](https://www.kaggle.com/models/atharvav75/dress-rehearsal-qwen3-kv4-grpo-adapters) |
| KV-cache precision study adapters (33 runs) | [corbanyax/KV-Cache-Precision-GRPO-Qwen3-0.6B-LoRA](https://huggingface.co/corbanyax/KV-Cache-Precision-GRPO-Qwen3-0.6B-LoRA) | [model](https://www.kaggle.com/models/atharvav75/kv-cache-precision-grpo-qwen3-lora-adapters) |
| Countdown-3 puzzles with hash-defined splits | [corbanyax/Countdown-3-Hash-Splits](https://huggingface.co/datasets/corbanyax/Countdown-3-Hash-Splits) | [dataset](https://www.kaggle.com/datasets/atharvav75/countdown-3-arithmetic-puzzles-hash-splits) |
| Full run logs, diagnostics, decision log | [corbanyax/KV-Cache-Precision-GRPO-Run-Logs](https://huggingface.co/datasets/corbanyax/KV-Cache-Precision-GRPO-Run-Logs) | [dataset](https://www.kaggle.com/datasets/atharvav75/dress-rehearsal-kv-precision-grpo-run-logs) |
Paper: *Beyond a Single Score: A Multi-Objective Evaluation of KV-Cache Quantization in GRPO Training* (under double-blind review, AAAI-27 Workshop on Multi-Objective and Multi-Task Optimization for Robust AI). All items are private until the review period ends.