open-thought/reasoning-gym
1
0
Fork 0
mirror of https://github.com/open-thought/reasoning-gym.git synced 2026-04-22 16:49:06 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Feat/intragen experiments (#414)

Browse source 
* added curriculum

* readapted readme

* corrected small errors

* Delete eval/eval/r1/algorithmic/word_sorting.json

* removed redundant argument

* added spell

* removed duplicated fit

* changed config

* added composite changes

* added composite changes

* updated yaml

* added spell backward

* updated read me

* added qwen2.5

* added

* Add files via upload

* updated missing trainer func

* updated curr

* updated spell back

* updated correctness score func

* updated configs

* added local evals

* added updates

* updated datasets

* added fsdp to hf utility

* added algorithmic qwen 3b yaml

* updated read me

* updated configs

* added preappend token

* updated with thinking token

* updated test score board

* resolved comments

* added evaluation scripts

* removed results from pr

* added config

* added partial reward scoring

* added evaluation composites

* added training configs

* added games eval

* added rubriks cube

* resolved merge cinflicts

* added games config

* added latest eval configs

* updated strucutre

* Delete training/evaluations/eval_graphs_composite.yaml

---------

Co-authored-by: joesharratt1229 <joesharrat1229@gmail.com>
This commit is contained in:
joesharratt1229 2025-04-16 07:04:52 +01:00 • committed by GitHub
parent 224532f12a
commit d0ef136d5b
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
21 changed files with 1331 additions and 48 deletions
Download patch file Download diff file Expand all files Collapse all files

42
reasoning_gym/cognition/rubiks_cube.py
View file

@ -121,29 +121,49 @@ class RubiksCubeDataset(ProceduralDataset):
},
}
def partial_score(self, cube: Cube) -> float:
"""
Returns a fraction between 0 and 1, indicating how many stickers are
correctly positioned (i.e., match the solved color for that face).
"""
total_stickers = 6 * (cube.size**2)
correct_stickers = 0
for face_index in range(6):
face = cube.faces[face_index]
solved_color = face[cube.size // 2][cube.size // 2].color
for row in range(cube.size):
for col in range(cube.size):
sticker = face[row][col]
if sticker.color == solved_color:
correct_stickers += 1
return correct_stickers / total_stickers
def score_answer(self, answer: Optional[str], entry: dict[str, Any]) -> float:
"""Determine if the solution provided solves the cube"""
reward = 0.0 # default reward
"""Determine if the solution provided solves the cube, with partial rewards."""
reward = 0.0 # default
if answer is not None:
# Reconstruct the test cube
eval_cube = Cube(entry["metadata"]["cube_size"])
eval_cube.rotate(entry["metadata"]["scramble_moves"])
# Test the solution
try:
expanded_answer = self.expand_moves(answer)
eval_cube.rotate(expanded_answer)
solved = eval_cube.is_done()
# 3) Check if fully solved
solved = eval_cube.is_done()
if solved:
reward = 1.0
elif len(answer.strip()) > 0: # encourage non-empty answers
reward = 0.05 # Incorrect, but rotate could parse the answer
else:
reward = 0.01
except:
reward = 0.01 # At least you tried
partial = self.partial_score(eval_cube)
if len(answer.strip()) > 0:
reward = max(0.05, partial)
else:
reward = max(0.01, partial)
except:
reward = 0.01
return reward
def remove_ansi(self, line):