reasoning-gym/training/evaluations/inter_generalisation/arc.yaml
Oliver Stanley 85f3c6dd02
updated inter-domain generalisation eval configs (#432)
* tweak eval configs

* add eval configs

* add eval config
2025-05-15 09:08:16 +02:00

33 lines
925 B
YAML

# Config used for evaluating inter-domain generalisation experiment models on ARC test data
# Models evaluated on this config:
# Qwen/Qwen2.5-3B-Instruct (original model)
# inter_algebra_qwen_3b_500 (original + 500 GRPO steps on algebra 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: arc
datasets:
- dataset: arc_1d
size: 100
seed: 42
- dataset: arc_agi
size: 100
seed: 42
- dataset: rearc
size: 100
seed: 42