SkillOpt v0.1.0: initial release

- Skill optimization framework with training loop analogy - 11 benchmarks, 4 model backends (Azure OpenAI, Claude, Codex, Qwen) - WebUI for browser-based training control - Pluggable architecture for extending benchmarks and backends
2026-05-21 17:22:04 +00:00
commit 244e346b83
237 changed files with 30248 additions and 0 deletions
@@ -0,0 +1,139 @@
+# Vendored from SkillRL (Apache-2.0 License)
+# Original: agent_system/environments/env_manager.py
+# Trimmed to only include AlfWorldEnvironmentManager and its helpers.
+
+from typing import List, Dict, Any
+from collections import defaultdict
+import numpy as np
+
+from skillopt.envs.alfworld.vendor.env_base import EnvironmentManagerBase, to_numpy
+from skillopt.envs.alfworld.vendor.alfworld_prompts import (
+    ALFWORLD_TEMPLATE,
+    ALFWORLD_TEMPLATE_NO_HIS,
+    ALFWORLD_TEMPLATE_WITH_MEMORY,
+)
+from skillopt.envs.alfworld.vendor.memory import SimpleMemory
+
+
+def parse_gamefile(infos):
+    gamefile = []
+    for info in infos:
+        if 'extra.gamefile' in info:
+            gamefile.append(info['extra.gamefile'])
+        else:
+            gamefile.append(None)
+    return gamefile
+
+
+def set_gamefile(infos, gamefile):
+    for i in range(len(infos)):
+        if 'extra.gamefile' in infos[i]:
+            infos[i]['extra.gamefile'] = gamefile[i]
+        else:
+            infos[i]['extra.gamefile'] = None
+    return infos
+
+
+class AlfWorldEnvironmentManager(EnvironmentManagerBase):
+    """Manages parallel ALFWorld environments with observation templating."""
+
+    def __init__(self, envs, projection_f, config):
+        self.memory = SimpleMemory()
+        self.retrieval_memory = None
+        super().__init__(envs, projection_f, config)
+
+    def reset(self, kwargs):
+        text_obs, image_obs, infos = self.envs.reset()
+        self.gamefile = parse_gamefile(infos)
+        self.memory.reset(batch_size=len(text_obs))
+        self.tasks = []
+        self.pre_text_obs = text_obs
+        self.extract_task(text_obs)
+
+        full_text_obs = self.build_text_obs(text_obs, self.envs.get_admissible_commands, init=True)
+        return {'text': full_text_obs, 'image': image_obs, 'anchor': text_obs}, infos
+
+    def step(self, text_actions: List[str]):
+        actions, valids = self.projection_f(text_actions, self.envs.get_admissible_commands)
+        text_obs, image_obs, rewards, dones, infos = self.envs.step(actions)
+        self.memory.store({'text_obs': self.pre_text_obs, 'action': actions})
+        self.pre_text_obs = text_obs
+
+        full_text_obs = self.build_text_obs(text_obs, self.envs.get_admissible_commands)
+        if infos[0].get("extra.gamefile") is None:
+            infos = set_gamefile(infos, self.gamefile)
+
+        for i, info in enumerate(infos):
+            info['is_action_valid'] = to_numpy(valids[i])
+
+        next_observations = {'text': full_text_obs, 'image': image_obs, 'anchor': text_obs}
+        rewards = to_numpy(rewards)
+        dones = to_numpy(dones)
+        return next_observations, rewards, dones, infos
+
+    def extract_task(self, text_obs: List[str]):
+        for obs in text_obs:
+            task_start = obs.find('Your task is to: ')
+            if task_start != -1:
+                self.tasks.append(obs[task_start + len('Your task is to: '):].strip())
+            else:
+                raise ValueError("Task description not found in text observation.")
+
+    def build_text_obs(self, text_obs: List[str], admissible_actions: List[List[str]], init: bool = False) -> List[str]:
+        postprocess_text_obs = []
+        if not init and self.config.env.history_length > 0:
+            memory_contexts, valid_lens = self.memory.fetch(
+                self.config.env.history_length,
+                obs_key="text_obs",
+                action_key="action",
+            )
+
+        for i in range(len(text_obs)):
+            reformatted_admissible_actions = "\n ".join(
+                f"'{s}'" for s in admissible_actions[i] if s != 'help'
+            )
+
+            if init or self.config.env.history_length <= 0:
+                obs = ALFWORLD_TEMPLATE_NO_HIS.format(
+                    current_observation=text_obs[i],
+                    admissible_actions=reformatted_admissible_actions,
+                )
+            else:
+                obs = ALFWORLD_TEMPLATE.format(
+                    task_description=self.tasks[i],
+                    step_count=len(self.memory[i]),
+                    history_length=valid_lens[i],
+                    action_history=memory_contexts[i],
+                    current_step=len(self.memory[i]) + 1,
+                    current_observation=text_obs[i],
+                    admissible_actions=reformatted_admissible_actions,
+                )
+            postprocess_text_obs.append(obs)
+        return postprocess_text_obs
+
+    def _process_batch(self, batch_idx, total_batch_list, total_infos, success):
+        for i in reversed(range(len(total_batch_list[batch_idx]))):
+            batch_item = total_batch_list[batch_idx][i]
+            if batch_item['active_masks']:
+                info = total_infos[batch_idx][i]
+                won_value = float(info['won'])
+                success['success_rate'].append(won_value)
+
+                gamefile = info.get("extra.gamefile")
+                if gamefile:
+                    self._process_gamefile(gamefile, won_value, success)
+                return
+
+    def _process_gamefile(self, gamefile, won_value, success):
+        tasks = [
+            "pick_and_place",
+            "pick_two_obj_and_place",
+            "look_at_obj_in_light",
+            "pick_heat_then_place_in_recep",
+            "pick_cool_then_place_in_recep",
+            "pick_clean_then_place_in_recep",
+        ]
+        for task in tasks:
+            if task in gamefile:
+                success[f"{task}_success_rate"].append(won_value)
+                break