Spaces:
Sleeping
Sleeping
File size: 11,540 Bytes
cc6473a ecbe0d8 cc6473a ecbe0d8 ece0bbe ecbe0d8 cc6473a ecbe0d8 cc6473a ecbe0d8 cc6473a ecbe0d8 ece0bbe cc6473a ecbe0d8 cc6473a ecbe0d8 cc6473a ecbe0d8 ece0bbe ecbe0d8 b9c9b8f ecbe0d8 cc6473a ecbe0d8 cc6473a ecbe0d8 cc6473a ecbe0d8 cc6473a ecbe0d8 ece0bbe cc6473a ecbe0d8 cc6473a ecbe0d8 cc6473a ecbe0d8 cc6473a ecbe0d8 cc6473a ecbe0d8 cc6473a ecbe0d8 cc6473a ecbe0d8 cc6473a ecbe0d8 cc6473a ecbe0d8 cc6473a ecbe0d8 cc6473a 1217c1d cc6473a 1217c1d cc6473a ece0bbe ecbe0d8 ece0bbe ecbe0d8 ece0bbe ecbe0d8 ece0bbe cc6473a ecbe0d8 cc6473a ecbe0d8 cc6473a ece0bbe ecbe0d8 cc6473a ecbe0d8 cc6473a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 | """
RhythmEnv Inference Evaluation — Baseline vs Trained, with meta-RL eval suite.
Three evaluation conditions:
1. discrete-3-profiles: 3 hardcoded reference profiles. A sanity check
that the meta-trained agent still handles the original named profiles.
2. continuous-in-distribution: Sampled profiles from the training distribution
(was the agent able to learn the meta-policy?)
3. continuous-OOD: Profiles from a held-out region of the parameter space
(does the meta-policy generalize, or did the agent memorize?)
Usage:
# Baselines only (no trained model):
python training/inference_eval.py
# With trained model:
python training/inference_eval.py --model_path outputs/rhythmenv_trained
"""
import argparse
import json
import os
import random
import sys
from typing import Optional
sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
from models import ActionType, RhythmAction
from server.rhythm_environment import RhythmEnvironment, MAX_STEPS, sample_profile, profile_to_belief_vector
from training.dataset import heuristic_action
DISCRETE_PROFILES = ["introvert_morning", "extrovert_night_owl", "workaholic_stoic"]
SLOT_NAMES = ["Morning", "Afternoon", "Evening", "Night"]
DAY_NAMES = ["Mon", "Tue", "Wed", "Thu", "Fri", "Sat", "Sun"]
# Seed ranges: training distribution = [0, 200); OOD = [10000, 10030)
# (10000 offset makes seeded sampled profiles in OOD region statistically distinct)
IN_DIST_SEEDS_DEFAULT = list(range(100, 110)) # 10 unseen-by-training in-distribution
OOD_SEEDS_DEFAULT = list(range(10000, 10010)) # 10 OOD seeds
def random_action(rng) -> ActionType:
return rng.choice(list(ActionType))
def model_action(obs, model, tokenizer, return_belief: bool = False):
"""Get action (and optionally belief) from trained model."""
# Lazy imports: keep the heavy training-stack imports out of module load
# so this script can run in baseline-only mode without unsloth/transformers.
from training.dataset import format_observation_prompt, SYSTEM_PROMPT
from training.reward_functions import extract_action_and_belief
prompt = format_observation_prompt(obs)
messages = [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": prompt},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
# 256 tokens lets the SFT-distilled student emit its full
# <reasoning>...</reasoning> block PLUS the final S M W ACTION_NAME line.
# Earlier 20-token cap truncated mid-reasoning so the answer line was
# never reached and parser fell back to extracting action names from
# the partial reasoning text.
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.7, do_sample=True)
response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
action_type, belief, _ = extract_action_and_belief(response)
if action_type is None:
action_type = ActionType.SLEEP
return (action_type, belief) if return_belief else action_type
def run_episode(
seed: int,
strategy: str,
profile_mode: str = "continuous",
profile: Optional[str] = None,
model=None,
tokenizer=None,
) -> dict:
"""Run a single episode and return per-episode metrics."""
rng = random.Random(seed + 500)
env = RhythmEnvironment()
if profile is not None:
obs = env.reset(seed=seed, profile=profile)
else:
obs = env.reset(seed=seed, profile_mode=profile_mode)
true_belief = env.get_belief_target()
profile_name = env.state.profile_name
total_reward = 0.0
step_rewards = []
actions_taken = []
beliefs_seen = [] # for trained model
for step in range(MAX_STEPS):
if obs.done:
break
if strategy == "heuristic":
action_type = heuristic_action(obs)
elif strategy == "random":
action_type = random_action(rng)
elif strategy == "model" and model is not None:
action_type, belief = model_action(obs, model, tokenizer, return_belief=True)
beliefs_seen.append(belief)
# Tell the env about the emitted belief so the grader can score
# belief_accuracy. Heuristic / random skip this — they get 0 on
# the belief component, by design.
env.record_belief(belief)
else:
action_type = random_action(rng)
action = RhythmAction(action_type=action_type)
actions_taken.append(action_type.value)
obs = env.step(action)
total_reward += obs.reward
step_rewards.append(obs.reward)
final_score = obs.reward_breakdown.get("final_score", 0.0)
# Adaptation: late-half mean minus early-half mean
half = max(len(step_rewards) // 2, 1)
early = step_rewards[:half]
late = step_rewards[half:]
adaptation = (sum(late) / len(late) - sum(early) / len(early)) if (early and late) else 0.0
# Belief tracking (only for trained model)
final_belief = beliefs_seen[-1] if beliefs_seen else None
belief_mae = None
if final_belief is not None:
belief_mae = sum(abs(b - t) for b, t in zip(final_belief, true_belief)) / 3.0
return {
"profile_name": profile_name,
"profile_mode": profile_mode if profile is None else "discrete",
"strategy": strategy,
"seed": seed,
"final_score": round(final_score, 4),
"total_reward": round(total_reward, 2),
"adaptation": round(adaptation, 3),
"vitality": round(obs.vitality, 2),
"cognition": round(obs.cognition, 2),
"progress": round(obs.progress, 2),
"serenity": round(obs.serenity, 2),
"connection": round(obs.connection, 2),
"actions": actions_taken,
"true_belief": [round(x, 3) for x in true_belief],
"final_belief": [round(x, 3) for x in final_belief] if final_belief is not None else None,
"belief_mae": round(belief_mae, 3) if belief_mae is not None else None,
}
def eval_condition(
name: str,
strategies: list[str],
runs: list[dict],
model=None,
tokenizer=None,
) -> list[dict]:
"""Run an eval condition and print summary."""
print(f"\n{'=' * 60}")
print(f"Condition: {name}")
print(f"{'=' * 60}")
results = []
for strategy in strategies:
print(f"\n Strategy: {strategy.upper()}")
scores = []
adaptations = []
belief_maes = []
for run in runs:
r = run_episode(strategy=strategy, model=model, tokenizer=tokenizer, **run)
results.append({"condition": name, **r})
scores.append(r["final_score"])
adaptations.append(r["adaptation"])
if r["belief_mae"] is not None:
belief_maes.append(r["belief_mae"])
avg_score = sum(scores) / len(scores) if scores else 0.0
avg_adapt = sum(adaptations) / len(adaptations) if adaptations else 0.0
avg_mae = sum(belief_maes) / len(belief_maes) if belief_maes else None
line = f" avg_score={avg_score:.3f} avg_adaptation={avg_adapt:+.3f}"
if avg_mae is not None:
line += f" avg_belief_mae={avg_mae:.3f}"
print(line)
return results
def main():
parser = argparse.ArgumentParser(description="Evaluate RhythmEnv agent (meta-RL eval suite)")
parser.add_argument("--model_path", type=str, default=None,
help="Path to trained model (skip for baseline only)")
parser.add_argument("--num_episodes", type=int, default=5,
help="Episodes per condition per strategy (for discrete: per-profile)")
parser.add_argument("--output_file", type=str, default="eval_results.json")
parser.add_argument("--in_dist_seeds", type=str, default=None,
help="Comma-separated seeds for in-distribution eval")
parser.add_argument("--ood_seeds", type=str, default=None,
help="Comma-separated seeds for OOD eval")
args = parser.parse_args()
in_dist_seeds = (
[int(s) for s in args.in_dist_seeds.split(",")] if args.in_dist_seeds
else IN_DIST_SEEDS_DEFAULT[:args.num_episodes * 2]
)
ood_seeds = (
[int(s) for s in args.ood_seeds.split(",")] if args.ood_seeds
else OOD_SEEDS_DEFAULT[:args.num_episodes * 2]
)
model, tokenizer = None, None
strategies = ["heuristic", "random"]
if args.model_path and os.path.exists(args.model_path):
try:
from unsloth import FastLanguageModel
# max_seq_length=2048 must accommodate: user prompt with 7-step
# history + per-meter anomalies (~900-1200 tokens) PLUS
# max_new_tokens=256 for the CoT response. Earlier value of 768
# silently truncated prompts on the LEFT (kept end of prompt,
# lost system instructions or older meter history), producing
# incoherent model outputs.
model, tokenizer = FastLanguageModel.from_pretrained(
model_name=args.model_path,
load_in_4bit=True,
max_seq_length=2048,
)
FastLanguageModel.for_inference(model)
strategies.append("model")
print(f"Loaded trained model from: {args.model_path}")
except Exception as e:
print(f"Warning: Could not load model: {e}")
print("Running baseline-only evaluation.")
all_results = []
# Condition 1: 3 hardcoded reference profiles (sanity-check the agent
# still handles the named profiles — no longer the primary eval signal).
discrete_runs = [
{"seed": ep, "profile": p}
for p in DISCRETE_PROFILES for ep in range(args.num_episodes)
]
all_results += eval_condition(
"discrete-3-profiles",
strategies, discrete_runs,
model=model, tokenizer=tokenizer,
)
# Condition 2: In-distribution sampled profiles
in_dist_runs = [{"seed": s, "profile_mode": "continuous"} for s in in_dist_seeds]
all_results += eval_condition(
"continuous-in-distribution",
strategies, in_dist_runs,
model=model, tokenizer=tokenizer,
)
# Condition 3: OOD sampled profiles (the meta-learning generalization test)
ood_runs = [{"seed": s, "profile_mode": "continuous"} for s in ood_seeds]
all_results += eval_condition(
"continuous-OOD (generalization)",
strategies, ood_runs,
model=model, tokenizer=tokenizer,
)
# Per-profile breakdown for the 3 reference profiles
print(f"\n{'=' * 70}")
print("DISCRETE-3-PROFILE BREAKDOWN")
print(f"{'=' * 70}")
print(f"{'Profile':<25} ", end="")
for s in strategies:
print(f"{s:>10}", end="")
print()
print("-" * 70)
discrete = [r for r in all_results if r["condition"] == "discrete-3-profiles"]
for profile in DISCRETE_PROFILES:
row = f"{profile:<25} "
for s in strategies:
rs = [r for r in discrete if r["profile_name"] == profile and r["strategy"] == s]
avg = sum(r["final_score"] for r in rs) / len(rs) if rs else 0.0
row += f"{avg:>10.3f}"
print(row)
# Save
with open(args.output_file, "w") as f:
json.dump(all_results, f, indent=2)
print(f"\nResults saved to: {args.output_file}")
if __name__ == "__main__":
main()
|