Spaces:
Running
Running
File size: 6,507 Bytes
df97e68 | 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 | """
scripts/convert_grpo_csv.py
Converts GRPO training CSV logs to JSON format
for the FastAPI /training/* story endpoints.
CSV format expected:
step, reward, fn1_valid, fn2_no_halluc, fn3_env_score
Usage:
python scripts/convert_grpo_csv.py \
--csv grpo_training_log.csv \
--task mixed_urgency_medium
Output:
data/training_logs/{task_id}_training_log.json
"""
from __future__ import annotations
import csv
import json
import argparse
from pathlib import Path
def load_csv(csv_path: str) -> list[dict]:
rows = []
with open(csv_path, newline="", encoding="utf-8") as f:
reader = csv.DictReader(f)
fieldnames = set(reader.fieldnames or [])
reward_values: list[float] = []
raw_rows: list[dict] = []
def _pick(row: dict, names: list[str], default: float) -> float:
for name in names:
if name in row and str(row.get(name, "")).strip() != "":
try:
return float(row[name])
except (TypeError, ValueError):
continue
return float(default)
for row in reader:
raw_rows.append(row)
reward_values.append(_pick(row, ["reward", "total_reward"], 0.0))
r_min = min(reward_values) if reward_values else 0.0
r_rng = (max(reward_values) - r_min) if reward_values else 1.0
if r_rng == 0:
r_rng = 1.0
for i, row in enumerate(raw_rows):
reward_val = _pick(row, ["reward", "total_reward"], 0.0)
fallback_norm = (reward_val - r_min) / r_rng
step_default = i + 1
if "step" in row and str(row.get("step", "")).strip() != "":
try:
step_default = int(float(row["step"]))
except (TypeError, ValueError):
step_default = i + 1
rows.append({
"step": step_default,
"reward": reward_val,
"fn1_valid": _pick(row, ["fn1_valid", "valid_action_rate"], 1.0),
"fn2_no_halluc": _pick(row, ["fn2_no_halluc", "hallucination_free"], 1.0),
"fn3_env_score": _pick(row, ["fn3_env_score", "env_score"], fallback_norm),
})
return rows
def build_log(rows: list[dict], task_id: str) -> dict:
n = len(rows)
rewards = [r["reward"] for r in rows]
fn1_vals = [r["fn1_valid"] for r in rows]
fn2_vals = [r["fn2_no_halluc"] for r in rows]
fn3_vals = [r["fn3_env_score"] for r in rows]
fn3_min = min(fn3_vals)
fn3_rng = (max(fn3_vals) - fn3_min) or 1.0
episodes = []
for i, r in enumerate(rows):
norm_env = (r["fn3_env_score"] - fn3_min) / fn3_rng
combined = round(
r["fn1_valid"] * 0.3 + r["fn2_no_halluc"] * 0.2 + norm_env * 0.5,
4
)
phase = (
"random" if i < n * 0.25 else
"exploring" if i < n * 0.50 else
"learning" if i < n * 0.75 else
"converged"
)
episodes.append({
"episode": r["step"],
"total_reward": round(r["reward"], 4),
"score": combined,
"fn1_valid": round(r["fn1_valid"], 4),
"fn2_no_halluc": round(r["fn2_no_halluc"], 4),
"fn3_env_score": round(r["fn3_env_score"], 4),
"phase": phase,
"actions": {
"valid_action_rate": round(r["fn1_valid"], 4),
"hallucination_free": round(r["fn2_no_halluc"], 4),
"env_score": round(norm_env, 4),
},
})
scores = [e["score"] for e in episodes]
return {
"task_id": task_id,
"total_episodes": n,
"base_model": "Qwen/Qwen2-1.5B-Instruct",
"adapter_path": f"artifacts/llm/{task_id.split('_')[1]}/",
"training_method": "GRPO",
"lora_rank": 16,
"reward_functions": {
"fn1_valid": "Action validity - legal JSON action output (0->1)",
"fn2_no_halluc": "No hallucination - stayed on gov workflow topic (0->1)",
"fn3_env_score": "Environment score - improved gov workflow quality",
},
"summary": {
"first_episode_reward": round(rewards[0], 4),
"last_episode_reward": round(rewards[-1], 4),
"best_episode_reward": round(max(rewards), 4),
"first_episode_score": round(scores[0], 4),
"last_episode_score": round(scores[-1], 4),
"best_episode_score": round(max(scores), 4),
"reward_improvement_pct": round(
((rewards[-1] - rewards[0]) / abs(rewards[0])) * 100, 2
) if rewards[0] != 0 else 0.0,
"invalid_action_steps": sum(1 for r in rows if r["fn1_valid"] < 1.0),
"hallucination_steps": sum(1 for r in rows if r["fn2_no_halluc"] < 1.0),
"avg_fn1_valid": round(sum(fn1_vals) / n, 4),
"avg_fn2_no_halluc": round(sum(fn2_vals) / n, 4),
"avg_fn3_env_score": round(sum(fn3_vals) / n, 4),
},
"episodes": episodes,
}
def save_log(log: dict, out_dir: str) -> str:
Path(out_dir).mkdir(parents=True, exist_ok=True)
out_path = f"{out_dir}/{log['task_id']}_training_log.json"
with open(out_path, "w", encoding="utf-8") as f:
json.dump(log, f, indent=2)
return out_path
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--csv", required=True, help="Path to GRPO CSV file")
parser.add_argument("--task", required=True, help="Task ID e.g. mixed_urgency_medium")
parser.add_argument("--out", default="data/training_logs", help="Output directory")
args = parser.parse_args()
print(f"Reading CSV : {args.csv}")
rows = load_csv(args.csv)
print(f"Steps found : {len(rows)}")
log = build_log(rows, args.task)
out = save_log(log, args.out)
print(f"Saved JSON : {out}")
print(f"Steps : {log['total_episodes']}")
print(f"Reward range : {log['summary']['first_episode_reward']} -> {log['summary']['last_episode_reward']}")
print(f"Score range : {log['summary']['first_episode_score']} -> {log['summary']['last_episode_score']}")
print(f"Invalid steps: {log['summary']['invalid_action_steps']}")
|