mirror of
https://github.com/open-thought/reasoning-gym.git
synced 2026-04-19 12:58:07 +00:00
45 lines
1.2 KiB
YAML
45 lines
1.2 KiB
YAML
# Config used for evaluating inter-domain generalisation experiment models on cognition test data
|
|
|
|
# Models evaluated on this config:
|
|
# Qwen/Qwen2.5-3B-Instruct (original model)
|
|
# inter_logic_qwen_3b_400 (original + 400 GRPO steps on logic RG data)
|
|
# inter_games_qwen_3b_500 (original + 500 GRPO steps on games RG data)
|
|
|
|
model_path: ../models/inter_games_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: cognition
|
|
datasets:
|
|
- dataset: color_cube_rotation
|
|
size: 100
|
|
seed: 42
|
|
- dataset: figlet_font
|
|
size: 100
|
|
seed: 42
|
|
- dataset: modulo_grid
|
|
size: 100
|
|
seed: 42
|
|
- dataset: needle_haystack
|
|
size: 100
|
|
seed: 42
|
|
- dataset: number_sequence
|
|
size: 100
|
|
seed: 42
|
|
- dataset: rectangle_count
|
|
size: 100
|
|
seed: 42
|
|
- dataset: rubiks_cube
|
|
size: 100
|
|
seed: 42
|