mirror of
https://github.com/open-thought/reasoning-gym.git
synced 2026-04-19 12:58:07 +00:00
10863ea12b
* add inter-domain generalisation eval config for algebra * add algorithmic eval cfg * vllm infer * add arithmetic eval cfg * add geometry eval cfg * add arc cfg * add games eval cfg * add cognition eval cfg * add graphs eval cfg
68 lines
1.7 KiB
YAML
68 lines
1.7 KiB
YAML
# Config used for evaluating inter-domain generalisation experiment models on arithmetic test data
|
|
|
|
# Models evaluated on this config:
|
|
# Qwen/Qwen2.5-3B-Instruct (original model)
|
|
# inter_algorithmic_qwen_3b_500 (original + 500 GRPO steps on algorithmic RG data)
|
|
|
|
model_path: ../models/inter_algorithmic_qwen_3b_500 # Change to the model to be evaluated
|
|
|
|
max_tokens: 2048 # From max_response_length in training config
|
|
top_p: 0.9 # From rollout top_p
|
|
temperature: 0.6 # Lower temperature for more focused responses
|
|
|
|
developer_prompt: DeepSeekZero
|
|
developer_role: system
|
|
|
|
output_dir: results
|
|
save_metadata: true
|
|
save_full_results: true
|
|
eval_repeats: 3
|
|
|
|
categories:
|
|
- category: arithmetic
|
|
datasets:
|
|
- dataset: basic_arithmetic
|
|
size: 100
|
|
seed: 42
|
|
- dataset: bitwise_arithmetic
|
|
size: 100
|
|
seed: 42
|
|
- dataset: calendar_arithmetic
|
|
size: 100
|
|
seed: 42
|
|
- dataset: chain_sum
|
|
size: 100
|
|
seed: 42
|
|
- dataset: count_bits
|
|
size: 100
|
|
seed: 42
|
|
- dataset: decimal_arithmetic
|
|
size: 100
|
|
seed: 42
|
|
- dataset: decimal_chain_sum
|
|
size: 100
|
|
seed: 42
|
|
- dataset: dice
|
|
size: 100
|
|
seed: 42
|
|
- dataset: fraction_simplification
|
|
size: 100
|
|
seed: 42
|
|
- dataset: gcd
|
|
size: 100
|
|
seed: 42
|
|
- dataset: lcm
|
|
size: 100
|
|
seed: 42
|
|
- dataset: power_function
|
|
size: 100
|
|
seed: 42
|
|
- dataset: prime_factorization
|
|
size: 100
|
|
seed: 42
|
|
- dataset: products
|
|
size: 100
|
|
seed: 42
|
|
- dataset: time_intervals
|
|
size: 100
|
|
seed: 42
|