ymh233's picture
Add files using upload-large-folder tool
8b97eb8 verified
Raw
History Blame Contribute Delete
8.93 kB
"""Shared helpers for working with saved trials under `results/`.
Consolidates the reusable pieces that used to live in the internal
`_count_tokens.py` / `_estimate_cost.py` / `_view_trial.py` scratch scripts:
- `iter_trials` walk results/<model>/<task>/<mode>/<exp>/trial_*.json
- `tokenize_chat` re-tokenise a chat_history with tiktoken
- `estimate_tokens_from_chat` prompt/completion split when usage_total is absent
- `PRICES` / `cost_from_usage` / `estimate_cost_from_chat` USD cost estimation
- `alias_to_full` model alias → exact provider model id
- `print_trial` pretty-print one saved trial JSON
Trial JSON layout (written by run_trial.py):
results/<model>/<task_id>/<mode>/<experiment>/trial_<seed:04d>_<ts>.json
{"meta": {...}, "trial": {... "usage_total": {...}}, "chat_history": [...],
"test_eval": {...}}
"""
from __future__ import annotations
import json
from pathlib import Path
from typing import Any, Dict, Iterator, List, Optional, Tuple
DEFAULT_RESULTS_DIR = Path(__file__).resolve().parent / "results"
# ---- trial iteration --------------------------------------------------------
def iter_trials(
results_dir: str | Path = DEFAULT_RESULTS_DIR,
model: Optional[str] = None,
) -> Iterator[Tuple[Path, Dict[str, Any], str]]:
"""Yield (path, parsed_json, model_alias) for every saved trial.
Walks `results/<model>/<task>/<mode>/<experiment>/trial_*.json`. When
`model` is given, restricts to that model's subtree. Files that fail to
parse, or that don't sit at the expected 5-part depth, are skipped.
"""
root = Path(results_dir)
glob_pat = (f"{model}/" if model else "*/") + "*/*/*/trial_*.json"
for trial_path in root.glob(glob_pat):
try:
data = json.loads(trial_path.read_text())
except Exception:
continue
parts = trial_path.relative_to(root).parts
if len(parts) < 5:
continue
yield trial_path, data, parts[0]
# ---- token counting ---------------------------------------------------------
def _get_encoder():
"""tiktoken encoder (o200k_base, falling back to cl100k_base), or None."""
try:
import tiktoken
except ImportError:
return None
try:
return tiktoken.get_encoding("o200k_base")
except Exception:
return tiktoken.get_encoding("cl100k_base")
def tokenize_chat(chat: List[Dict[str, Any]]) -> Optional[int]:
"""Total token count of every message's content, or None if tiktoken
is unavailable."""
enc = _get_encoder()
if enc is None:
return None
return sum(len(enc.encode(m.get("content") or "")) for m in chat)
def estimate_tokens_from_chat(chat: List[Dict[str, Any]]) -> Optional[Dict[str, int]]:
"""Approximate prompt/completion split when usage_total is missing.
Each assistant turn is one LLM call: prompt_tokens at that turn = tokens
of all prior messages; completion_tokens = tokens of the assistant
message. Returns None if tiktoken is unavailable.
"""
enc = _get_encoder()
if enc is None:
return None
def n(s: str) -> int:
return len(enc.encode(s or ""))
prompt_total = 0
completion_total = 0
running_prompt = 0
for msg in chat:
toks = n(msg.get("content", ""))
if msg.get("role") == "assistant":
prompt_total += running_prompt
completion_total += toks
running_prompt += toks # this turn becomes part of the next prompt
else:
running_prompt += toks
return {
"prompt_tokens": prompt_total,
"completion_tokens": completion_total,
"prompt_cached_tokens": 0,
"reasoning_tokens": 0,
}
# ---- cost estimation --------------------------------------------------------
# USD per 1M tokens (regular tier; off-peak DeepSeek = 50% off but we use full
# price as an upper bound). Update if rates change. Keyed by exact model id.
PRICES: Dict[str, Dict[str, float]] = {
# OpenAI
"gpt-4.1-mini-2025-04-14": {"in": 0.40, "in_cached": 0.10, "out": 1.60},
"gpt-4.1-2025-04-14": {"in": 2.00, "in_cached": 0.50, "out": 8.00},
"o4-mini-2025-04-16": {"in": 1.10, "in_cached": 0.275, "out": 4.40},
"gpt-5-nano": {"in": 0.05, "in_cached": 0.005, "out": 0.40},
"gpt-5-mini-2025-08-07": {"in": 0.25, "in_cached": 0.025, "out": 2.00},
"gpt-5": {"in": 1.25, "in_cached": 0.125, "out": 10.00},
"gpt-5-pro": {"in": 15.0, "in_cached": 1.50, "out": 120.0},
"gpt-5-codex": {"in": 1.25, "in_cached": 0.125, "out": 10.00},
"gpt-5.1": {"in": 1.25, "in_cached": 0.125, "out": 10.00},
"gpt-5.4": {"in": 1.25, "in_cached": 0.125, "out": 10.00},
"gpt-5.5": {"in": 1.25, "in_cached": 0.125, "out": 10.00},
"gpt-5.5-pro": {"in": 15.0, "in_cached": 1.50, "out": 120.0},
# DeepSeek
"deepseek-chat": {"in": 0.27, "in_cached": 0.07, "out": 1.10},
"deepseek-reasoner": {"in": 0.55, "in_cached": 0.14, "out": 2.19},
"deepseek-v4-flash": {"in": 0.27, "in_cached": 0.07, "out": 1.10},
"deepseek-v4-pro": {"in": 0.55, "in_cached": 0.14, "out": 2.19},
}
def cost_from_usage(usage: Dict[str, Any], model: str) -> Optional[float]:
"""USD cost from an exact `usage_total` breakdown, or None if `model`
isn't in PRICES. `completion_tokens` already includes reasoning tokens
for OpenAI-style usage."""
rates = PRICES.get(model)
if rates is None:
return None
p_total = int(usage.get("prompt_tokens", 0))
p_cached = int(usage.get("prompt_cached_tokens", 0))
p_uncached = max(p_total - p_cached, 0)
c = int(usage.get("completion_tokens", 0))
return (p_uncached * rates["in"] + p_cached * rates["in_cached"]
+ c * rates["out"]) / 1e6
def estimate_cost_from_chat(
chat: List[Dict[str, Any]], model: str
) -> Tuple[Optional[float], Optional[Dict[str, Any]]]:
"""Approximate (usd, token_breakdown) by re-tokenising chat_history when
usage_total is missing. Returns (None, None) if tiktoken is unavailable
or `model` isn't priced."""
rates = PRICES.get(model)
if rates is None:
return None, None
est = estimate_tokens_from_chat(chat)
if est is None:
return None, None
cost = (est["prompt_tokens"] * rates["in"]
+ est["completion_tokens"] * rates["out"]) / 1e6
return cost, {**est, "estimated": True}
# ---- model aliases ----------------------------------------------------------
def alias_to_full(alias: str) -> str:
"""Resolve a model alias to its exact provider model id via
call_llm_api.api_source_mapping. Returns `alias` unchanged if unknown."""
try:
from call_llm_api import api_source_mapping
return api_source_mapping.get(alias, ("?", alias))[1]
except Exception:
return alias
# ---- trial viewer -----------------------------------------------------------
def print_trial(data: Dict[str, Any]) -> None:
"""Pretty-print one saved trial JSON (meta, test eval, full chat history)."""
trial = data["trial"]
print("=== TRIAL META ===")
print(f" status={trial['status']} rounds={trial['rounds']} "
f"tokens={trial.get('total_tokens')} elapsed={data.get('meta', {}).get('elapsed_s', data.get('elapsed_s', 0)):.1f}s")
print(f" python_calls={trial.get('n_python_calls', 0)} "
f"experiments={trial.get('n_experiments', 0)} "
f"data_requests={trial.get('n_data_requests', 0)} "
f"rows_seen={trial.get('n_unique_rows_seen', 0)}")
e = data.get("test_eval")
if e:
print("\n=== TEST EVAL ===")
if e["status"] == "ok":
print(f" smape={e['smape']:.4f} mae={e['mae']:.4f} rmse={e['rmse']:.4f}")
bid, bval, ds = e.get("best_baseline_id"), e.get("best_baseline_value"), e.get("discovery_score")
if bid is not None and bval is not None and ds is not None and ds == ds:
print(f" discovery_score={ds:.4f} vs baseline {bid}={bval:.4f}")
else:
print(" discovery_score=N/A (no baseline reference)")
else:
print(f" status={e['status']} error={e.get('error')}")
chat = data.get("chat_history") or []
print(f"\n=== CHAT HISTORY ({len(chat)} messages) ===")
for i, msg in enumerate(chat):
bar = "═" * 78
print(f"\n{bar}")
print(f"[{i}] role={msg['role']} len={len(msg['content'])}")
print(bar)
print(msg["content"])
if __name__ == "__main__":
import sys
if len(sys.argv) > 1:
print_trial(json.loads(Path(sys.argv[1]).read_text()))
else:
print("usage: python utils.py <trial.json> # pretty-print a saved trial")