mirror of
https://github.com/open-thought/reasoning-gym.git
synced 2026-04-23 16:55:05 +00:00
a969d8ef05
* configs * reduce complexity of curriculum * update lower bound * add failure threshold * update last_k * update thresholds for success and failure * update curriculum file as well * update run name for noncurriculum * lint * dtype model eval * return binary scoring * set eval repeats to 3 * fix tests
34 lines
1.1 KiB
YAML
34 lines
1.1 KiB
YAML
# Config used for evaluating curriculum generalisation experiment models on Knights and Knaves
|
|
|
|
# Models evaluated on this config:
|
|
# Qwen/Qwen2.5-3B-Instruct (original model)
|
|
# qwen3b_knights-knaves_noncurriculum (original + 300 GRPO steps on non-curriculum Knights and Knaves data)
|
|
# qwen3b_knights-knaves_curriculum (original + 300 GRPO steps on curriculum Knights and Knaves data)
|
|
|
|
model_path: Qwen/Qwen2.5-3B-Instruct # Default model path
|
|
# model_path: /workspace/reasoning-gym/training/qwen3b_knights-knaves_noncurriculum
|
|
# model_path: /workspace/reasoning-gym/training/qwen3b_knights-knaves_curriculum
|
|
|
|
max_tokens: 2048 # From max_response_length in training config
|
|
top_p: 1.0
|
|
temperature: 1.0 # Lower temperature for more focused responses
|
|
dtype: bfloat16
|
|
|
|
developer_prompt: DeepSeekZero
|
|
developer_role: system
|
|
|
|
output_dir: results
|
|
save_metadata: true
|
|
save_full_results: true
|
|
eval_repeats: 3
|
|
|
|
categories:
|
|
- category: logic
|
|
datasets:
|
|
- dataset: knights_knaves
|
|
size: 100
|
|
seed: 42
|
|
params:
|
|
n_people: 4
|
|
depth_constraint: 3
|
|
width_constraint: 3
|