VAGEN / utils /transfer_rubikscube.py
Harryis's picture
Upload folder using huggingface_hub
fcc38f9 verified
Raw
History Blame Contribute Delete
17.8 kB
#!/usr/bin/env python3
"""
Convert Rubik's Cube DQN trajectories to SFT JSON.
Prompt layout follows VAGEN_old/vagen/env/rubikscube/prompt.py (FORMAT_CONFIGS + templates).
By default, actions in <answer> and "Last valid action(s)" use natural language aligned with
the reasoning style in FORMAT_CONFIGS (e.g. "Rotate Up clockwise"); use --action_repr token
for canonical symbols (UpCW, UpCCW, ...).
"""
from __future__ import annotations
import argparse
import importlib.util
import json
from pathlib import Path
from typing import Any, Dict, List, Optional, Tuple
_REPO_ROOT = Path(__file__).resolve().parents[1]
def _load_prompt_module(rel_path: str) -> Any:
"""Load vagen/env/*/prompt.py without importing vagen.env package __init__ (avoids side effects)."""
path = _REPO_ROOT / rel_path
name = "vagen_prompt_" + rel_path.replace("/", "_").replace(".py", "")
spec = importlib.util.spec_from_file_location(name, path)
if spec is None or spec.loader is None:
raise ImportError(f"Cannot load prompt module from {path}")
mod = importlib.util.module_from_spec(spec)
spec.loader.exec_module(mod)
return mod
_rc_prompt = _load_prompt_module("vagen/env/rubikscube/prompt.py")
rc_action_template = _rc_prompt.action_template
rc_format_prompt = _rc_prompt.format_prompt
rc_init_observation_template = _rc_prompt.init_observation_template
rc_system_prompt = _rc_prompt.system_prompt
# Keep transfer robust if prompt.py changes/gets swapped.
# We only append this block when the base system prompt lacks the face-name mapping.
_NET_MAPPING_BLOCK = """\nVision observation (IMPORTANT):\n- The image shows the cube as a 2D unfolded net (a cross) with 6 faces, each face is a 2x2 grid.\n- Face names in the unfolded net are fixed as:\n\n [U]\n[L] [F] [R] [B]\n [D]\n\n Where:\n - U = Up (top face)\n - D = Down (bottom face)\n - F = Front (facing you)\n - B = Back (opposite of Front)\n - L = Left (left of Front)\n - R = Right (right of Front)\n\n- Sticker positions inside each 2x2 face in the image are:\n - index 0: top-left\n - index 1: top-right\n - index 2: bottom-left\n - index 3: bottom-right\n"""
# --- Same action order as visual_scout/dqn_rubikscube.VagenRubiksCubeVisionWrapper.ACTIONS ---
ACTION_ID_TO_WORD = {
0: "UpCW",
1: "UpCCW",
2: "DownCW",
3: "DownCCW",
4: "LeftCW",
5: "LeftCCW",
6: "RightCW",
7: "RightCCW",
8: "FrontCW",
9: "FrontCCW",
10: "BackCW",
11: "BackCCW",
}
PROMPT_FORMAT_CHOICES = tuple(rc_format_prompt.keys())
def _token_to_natural(token: str) -> str:
"""Natural phrasing consistent with rubikscube/prompt.py worldmodeling example ('Rotate Up clockwise.')."""
if token.endswith("CCW"):
face = token[:-3]
return f"Rotate {face} counter-clockwise"
if token.endswith("CW"):
face = token[:-2]
return f"Rotate {face} clockwise"
raise ValueError(f"Unexpected action token: {token}")
def _action_text(action_id: int, *, action_repr: str) -> str:
"""Text inside <answer> and in Last valid action(s)."""
if int(action_id) not in ACTION_ID_TO_WORD:
raise ValueError(f"Unknown action id: {action_id}")
token = ACTION_ID_TO_WORD[int(action_id)]
if action_repr == "token":
return token
if action_repr == "natural":
return _token_to_natural(token)
raise ValueError(f"Unknown action_repr: {action_repr}")
def build_system_text(prompt_format: str, max_actions_per_step: int, action_sep: str) -> str:
if prompt_format not in PROMPT_FORMAT_CHOICES:
raise ValueError(f"Unknown prompt_format: {prompt_format}, expected one of {PROMPT_FORMAT_CHOICES}")
fmt_block = rc_format_prompt[prompt_format](max_actions_per_step, action_sep, add_example=True)
base = rc_system_prompt()
# Append mapping block only if not already present (avoid duplication).
if "Face names in the unfolded net are fixed as" not in base:
base = base.rstrip() + "\n" + _NET_MAPPING_BLOCK.lstrip("\n")
return base + "\n" + fmt_block
def format_block_only(prompt_format: str, max_actions_per_step: int, action_sep: str) -> str:
return rc_format_prompt[prompt_format](max_actions_per_step, action_sep, add_example=False)
def _assistant_action_text(action_word: str, prompt_format: str, think: str = "") -> str:
if think:
return f"<think>{think}</think><answer>{action_word}</answer>"
if prompt_format == "free_think":
return f"<think> </think><answer>{action_word}</answer>"
if prompt_format == "grounding":
return (
f"<think><observation> </observation><reasoning> </reasoning></think>"
f"<answer>{action_word}</answer>"
)
if prompt_format == "worldmodeling":
return (
f"<think><reasoning> </reasoning><prediction> </prediction></think>"
f"<answer>{action_word}</answer>"
)
if prompt_format == "grounding_worldmodeling":
return (
f"<think><observation> </observation><reasoning> </reasoning><prediction> </prediction></think>"
f"<answer>{action_word}</answer>"
)
raise ValueError(f"Unhandled prompt_format: {prompt_format}")
def _split_text_by_placeholder(text: str, placeholder: str = "<image>") -> List[Dict[str, Any]]:
parts = text.split(placeholder)
if len(parts) == 1:
return [{"type": "text", "text": text}]
content: List[Dict[str, Any]] = []
for i, p in enumerate(parts):
if p:
content.append({"type": "text", "text": p})
if i < len(parts) - 1:
content.append({"type": "image"})
return content
def _fill_image_blocks(content: List[Dict[str, Any]], image_path: str) -> List[Dict[str, Any]]:
out: List[Dict[str, Any]] = []
for block in content:
if block.get("type") == "image" and "image" not in block:
out.append({"type": "image", "image": image_path})
else:
out.append(block)
return out
def _blocks_to_sharegpt_content_and_images(
blocks: List[Dict[str, Any]], image_placeholder: str = "<image>"
) -> Tuple[str, List[str]]:
parts: List[str] = []
images: List[str] = []
for b in blocks:
btype = b.get("type")
if btype == "text":
parts.append(str(b.get("text", "")))
elif btype == "image":
parts.append(image_placeholder)
img = b.get("image")
if img is not None:
images.append(str(img))
else:
parts.append(str(b))
return "".join(parts), images
def messages_to_llamafactory_sharegpt(
messages: List[Dict[str, Any]], *, image_placeholder: str = "<image>"
) -> Dict[str, Any]:
out_messages: List[Dict[str, Any]] = []
out_images: List[str] = []
for m in messages:
role = m.get("role")
content = m.get("content")
if isinstance(content, list):
text, imgs = _blocks_to_sharegpt_content_and_images(content, image_placeholder=image_placeholder)
out_messages.append({"role": role, "content": text})
out_images.extend(imgs)
else:
out_messages.append({"role": role, "content": "" if content is None else str(content)})
return {"messages": out_messages, "images": out_images}
def build_messages_for_episode(
frames: List[str],
actions: List[int],
rewards: Optional[List[float]] = None,
*,
action_repr: str,
prompt_format: str,
max_actions_per_step: int,
action_sep: str,
include_reward: bool,
assistant_think: str,
) -> List[Dict[str, Any]]:
if len(frames) != len(actions) + 1:
raise ValueError(f"Expected len(frames)=len(actions)+1, got {len(frames)} vs {len(actions)}")
sys_text = build_system_text(prompt_format, max_actions_per_step, action_sep)
user_format_suffix = format_block_only(prompt_format, max_actions_per_step, action_sep)
messages: List[Dict[str, Any]] = [{"role": "system", "content": sys_text}]
init_text = rc_init_observation_template("<image>") + "\n" + user_format_suffix
init_content = _fill_image_blocks(_split_text_by_placeholder(init_text), frames[0])
messages.append({"role": "user", "content": init_content})
for t, act_id in enumerate(actions):
act_word = _action_text(act_id, action_repr=action_repr)
messages.append(
{
"role": "assistant",
"content": _assistant_action_text(act_word, prompt_format, think=assistant_think),
}
)
obs_text = rc_action_template([act_word], "<image>") + "\n" + user_format_suffix
if include_reward:
r = 0.0
if rewards is not None and t < len(rewards):
try:
r = float(rewards[t])
except Exception:
r = 0.0
obs_text = f"Reward:\n{r}\n\n" + obs_text
obs_content = _fill_image_blocks(_split_text_by_placeholder(obs_text), frames[t + 1])
messages.append({"role": "user", "content": obs_content})
return messages
def extract_system_prefix(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
sys_msgs: List[Dict[str, Any]] = []
for m in messages:
if m.get("role") == "system":
sys_msgs.append(m)
else:
break
return sys_msgs
def collect_user_assistant_pairs(messages: List[Dict[str, Any]], start_idx: int = 0) -> List[List[Dict[str, Any]]]:
pairs: List[List[Dict[str, Any]]] = []
i = int(start_idx)
n = len(messages)
while i < n:
while i < n and messages[i].get("role") != "user":
i += 1
if i >= n:
break
j = i + 1
if j < n and messages[j].get("role") == "assistant":
pairs.append([messages[i], messages[j]])
i = j + 1
else:
i += 1
return pairs
def split_conversation_cumulative(messages: List[Dict[str, Any]], source_id: int) -> List[Dict[str, Any]]:
sys_prefix = extract_system_prefix(messages)
start_idx = len(sys_prefix)
pairs = collect_user_assistant_pairs(messages, start_idx=start_idx)
outputs: List[Dict[str, Any]] = []
total_turns = len(pairs)
for k in range(1, total_turns + 1):
out_msgs = sys_prefix + [m for pair in pairs[:k] for m in pair]
outputs.append(
{
"messages": out_msgs,
"meta": {
"source_id": int(source_id),
"turns": int(k),
"total_turns": int(total_turns),
},
}
)
return outputs
def find_latest_step_dir(traj_root: Path) -> Path:
step_dirs = [p for p in traj_root.iterdir() if p.is_dir() and p.name.startswith("step_")]
if not step_dirs:
raise FileNotFoundError(f"No step_* directories under {traj_root}")
step_dirs.sort(key=lambda p: int(p.name.split("_")[-1]))
return step_dirs[-1]
def convert_step_dir(
step_dir: Path,
output_dir: Path,
*,
include_failed: bool,
action_repr: str,
prompt_format: str,
max_actions_per_step: int,
action_sep: str,
include_reward: bool,
assistant_think: str,
strip_prefix: Optional[str],
split_multiturn: bool,
output_format: str,
) -> Path:
traj_path = step_dir / "trajectories.jsonl"
metrics_path = step_dir / "metrics.json"
if not traj_path.exists():
raise FileNotFoundError(f"Missing trajectories.jsonl at {traj_path}")
output_dir.mkdir(parents=True, exist_ok=True)
out_path = output_dir / f"{step_dir.name}_sft.json"
global_step = None
if metrics_path.exists():
try:
global_step = json.loads(metrics_path.read_text()).get("global_step")
except Exception:
global_step = None
prefix = str(Path(strip_prefix)) if strip_prefix else None
out_items: List[Dict[str, Any]] = []
source_id = 0
with traj_path.open("r", encoding="utf-8") as fin:
for line in fin:
line = line.strip()
if not line:
continue
traj = json.loads(line)
ep_success = bool(traj.get("episode_success", False))
if (not include_failed) and (not ep_success):
continue
frames = traj.get("frames", [])
actions = traj.get("actions", [])
rewards = traj.get("rewards", [])
if not frames or not actions:
continue
frames = [str(Path(p).resolve()) for p in frames]
if prefix:
new_frames = []
for p in frames:
ps = str(p)
if ps.startswith(prefix):
ps = ps[len(prefix) :]
if ps.startswith("/"):
ps = ps[1:]
new_frames.append(ps)
frames = new_frames
messages = build_messages_for_episode(
frames=frames,
actions=actions,
rewards=rewards,
action_repr=action_repr,
prompt_format=prompt_format,
max_actions_per_step=max_actions_per_step,
action_sep=action_sep,
include_reward=include_reward,
assistant_think=assistant_think,
)
meta: Dict[str, Any] = {
"episode_return": traj.get("episode_return", None),
"episode_success": ep_success,
"global_step": global_step,
}
if split_multiturn:
split_records = split_conversation_cumulative(messages, source_id=source_id)
for sr in split_records:
merged_meta = dict(meta)
merged_meta.update(sr.get("meta", {}))
if output_format == "llamafactory_sharegpt":
rec = messages_to_llamafactory_sharegpt(sr["messages"])
out_items.append({**rec, "meta": merged_meta})
else:
out_items.append({"messages": sr["messages"], "meta": merged_meta})
else:
if output_format == "llamafactory_sharegpt":
rec = messages_to_llamafactory_sharegpt(messages)
out_items.append({**rec, "meta": meta})
else:
out_items.append({"messages": messages, "meta": meta})
source_id += 1
out_path.write_text(json.dumps(out_items, ensure_ascii=False, indent=2), encoding="utf-8")
return out_path
def main():
parser = argparse.ArgumentParser(
description="Convert visual Rubik's Cube eval trajectories to SFT JSON (vagen/env/rubikscube/prompt.py)."
)
parser.add_argument("run_dir", help="Path to the run directory (contains trajectories/)")
parser.add_argument("--step", default=None, help="Specific step directory name (e.g., step_400000)")
parser.add_argument("--include_failed", action="store_true", help="Include failed episodes")
parser.add_argument(
"--action_repr",
type=str,
default="natural",
choices=["natural", "token"],
help='How to write each action in <answer> and "Last valid action(s)": '
'natural = "Rotate Up clockwise" style (matches rubikscube/prompt.py examples); '
"token = UpCW, UpCCW, ... (canonical symbols from system prompt).",
)
parser.add_argument(
"--prompt_format",
type=str,
default="grounding_worldmodeling",
choices=PROMPT_FORMAT_CHOICES,
help="Must match a key in rubikscube/prompt.py FORMAT_CONFIGS",
)
parser.add_argument("--max_actions_per_step", type=int, default=1, help="max_actions_per_step (format block)")
parser.add_argument("--action_sep", type=str, default=",", help="Action separator in multi-action examples")
parser.add_argument("--include_reward", action="store_true", help="Prefix each observation turn with reward")
parser.add_argument("--assistant_think", type=str, default="", help="Assistant </think> content (optional)")
parser.add_argument("--strip_prefix", type=str, default=None, help="Optional path prefix to strip from frame paths")
parser.add_argument(
"--output_format",
type=str,
default="llamafactory_sharegpt",
choices=["llamafactory_sharegpt", "internal_blocks"],
help="Output JSON format. Use llamafactory_sharegpt for LLaMAFactory (messages+images).",
)
parser.add_argument(
"--no_split_multiturn",
action="store_true",
help="Disable cumulative multi-turn splitting; output one sample per episode.",
)
args = parser.parse_args()
run_dir = Path(args.run_dir)
traj_root = run_dir / "trajectories"
if not traj_root.exists():
raise FileNotFoundError(f"Not found trajectories directory: {traj_root}")
step_dir = traj_root / args.step if args.step else find_latest_step_dir(traj_root)
output_dir = run_dir / "sft"
out_path = convert_step_dir(
step_dir=step_dir,
output_dir=output_dir,
include_failed=args.include_failed,
action_repr=args.action_repr,
prompt_format=args.prompt_format,
max_actions_per_step=args.max_actions_per_step,
action_sep=args.action_sep,
include_reward=args.include_reward,
assistant_think=args.assistant_think,
strip_prefix=args.strip_prefix,
split_multiturn=(not args.no_split_multiturn),
output_format=args.output_format,
)
print(f"SFT data written to: {out_path}")
if __name__ == "__main__":
main()