File size: 4,287 Bytes
e890160 aaad9f1 e890160 aaad9f1 e890160 aaad9f1 e890160 | 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 | """
eval.py — Evaluate base vs fine-tuned model on the OpenEnv.
Runs episodes with:
1. The fine-tuned model (current LoRA adapter)
2. The heuristic baseline
Compares average rewards across tasks. Results are saved locally and pushed
via push_run_files_to_hub (in train.py) under hub_model_repo/<run_id>/eval_results.json.
"""
from __future__ import annotations
import json
import time
from typing import Any, Dict, List
import torch
try:
from .model_utils import push_to_hub
from .openenv_loop import (
OpenEnvClient,
rollout_episode,
rollout_heuristic_episode,
)
except ImportError:
from model_utils import push_to_hub
from openenv_loop import (
OpenEnvClient,
rollout_episode,
rollout_heuristic_episode,
)
def evaluate(
client: OpenEnvClient,
model,
tokenizer,
cfg: Dict[str, Any],
output_dir: str = "/tmp/antiatropos_eval",
) -> Dict[str, Any]:
"""Run evaluation: fine-tuned model vs heuristic baseline.
Returns a dict with per-task results and overall comparison.
"""
tasks = cfg.get("tasks", ["task-1", "task-2", "task-3"])
eval_episodes = cfg.get("eval_episodes", 3)
eval_max_steps = cfg.get("eval_max_steps", 60)
# Enable inference mode
try:
from unsloth import FastLanguageModel
FastLanguageModel.for_inference(model)
except ImportError:
model.eval()
results: Dict[str, Any] = {}
all_ft_rewards: List[float] = []
all_heur_rewards: List[float] = []
print(f"\n{'='*70}")
print(f"EVALUATION — {eval_episodes} episodes per task, {eval_max_steps} steps")
print(f"{'='*70}")
for task_id in tasks:
ft_rewards: List[float] = []
heur_rewards: List[float] = []
ft_invalid = 0
for ep in range(eval_episodes):
seed = 1000 + ep # Deterministic eval seeds
# Fine-tuned model episode
ft_ep = rollout_episode(
client, model, tokenizer, task_id,
eval_max_steps, cfg, seed=seed,
)
ft_rewards.append(ft_ep.avg_reward)
ft_invalid += ft_ep.num_invalid
# Heuristic baseline episode
heur_ep = rollout_heuristic_episode(
client, task_id, eval_max_steps, seed=seed,
)
heur_rewards.append(heur_ep.avg_reward)
ft_avg = sum(ft_rewards) / len(ft_rewards)
heur_avg = sum(heur_rewards) / len(heur_rewards)
all_ft_rewards.extend(ft_rewards)
all_heur_rewards.extend(heur_rewards)
winner = "FT WINS" if ft_avg >= heur_avg else "HEURISTIC WINS"
results[task_id] = {
"ft_avg_reward": ft_avg,
"heuristic_avg_reward": heur_avg,
"ft_wins": ft_avg >= heur_avg,
"ft_invalid_actions": ft_invalid,
}
print(f"\n {task_id}:")
print(f" FT model avg reward: {ft_avg:.4f}")
print(f" Heuristic avg reward: {heur_avg:.4f}")
print(f" Result: {winner}")
print(f" Invalid actions (FT): {ft_invalid}")
# Overall summary
tasks_won = sum(1 for r in results.values() if r["ft_wins"])
ft_overall = sum(all_ft_rewards) / len(all_ft_rewards) if all_ft_rewards else 0
heur_overall = sum(all_heur_rewards) / len(all_heur_rewards) if all_heur_rewards else 0
summary = {
"per_task": results,
"overall_ft_avg": ft_overall,
"overall_heuristic_avg": heur_overall,
"tasks_won_by_ft": tasks_won,
"total_tasks": len(tasks),
"ft_overall_wins": ft_overall >= heur_overall,
}
print(f"\n{'='*70}")
print(f"EVALUATION SUMMARY")
print(f"{'='*70}")
print(f" FT model overall avg: {ft_overall:.4f}")
print(f" Heuristic overall avg: {heur_overall:.4f}")
print(f" FT wins on: {tasks_won}/{len(tasks)} tasks")
print(f" Overall: {'FT WINS' if ft_overall >= heur_overall else 'HEURISTIC WINS'}")
# Save eval results
import os
os.makedirs(output_dir, exist_ok=True)
eval_path = f"{output_dir}/eval_results.json"
with open(eval_path, "w") as f:
json.dump(summary, f, indent=2)
print(f" [eval] Saved results → {eval_path}")
return summary
|