Mechanics of Learned Reasoning 1: TempoBench

Formally grounded LLM benchmark for temporal reasoning over automata/traces. Runs locally via a clean CLI or Python API. Keeps the wheel thin (code + tiny samples) and lets you plug in any model (OpenAI-compatible, Hugging Face, vLLM, or your own class/function).

---

Features

• Tasks: trace acceptance & temporal causality (with per-feature metrics).
• Backends: OpenRouter/OpenAI (OpenAI-compatible), HF pipelines, vLLM, or custom Python adapters.
• Outputs: row-wise JSONL + CSV with accuracy and F1s (AP & timestep).
• Reproducible runs: fixed seeds, manifest-friendly outputs, small packaged sample datasets.

---

Datasets

Check my huggingface for the tempobench public benchmarking datasets.

If you are interested in access to our datasets for reasoning SFT, reach out to me.

---

Install

pip install tempobench

Python ≥ 3.10 recommended.

---

Quickstart (CLI)

# OpenRouter (OpenAI-compatible) example
export OPENROUTER_API_KEY=YOUR_KEY

tempobench run \
  --dataset_path src/tempobench/data/causal-done.jsonl \
  --task causal \
  --backend openrouter \
  --model-id openai/gpt-4o-mini \
  --gen-args '{"temperature":0.0,"max_tokens":256}' \
  --outdir benchmark_results --console-prints

Other backends: This is not implemented yet and is an open issue currently.

# OpenAI
export OPENAI_API_KEY=YOUR_KEY
tempobench run --dataset_path ... --task causal --backend openai --model-id gpt-4o-mini

# Hugging Face (local model)
tempobench run --dataset_path ... --task causal \
  --backend hf --model-id meta-llama/Meta-Llama-3.1-8B-Instruct \
  --model-args '{"device":0}' --gen-args '{"max_new_tokens":256}'

# vLLM server (OpenAI API compatible)
tempobench run --dataset_path ... --task trace \
  --backend vllm --model-id my-vllm \
  --model-args '{"base_url":"http://127.0.0.1:8000/v1","api_key":"nokey"}'

Outputs land under benchmark_results/<task>/ as both .jsonl and .csv.

---

Python API

You can use the benchmarker to build custom benchmarking workflows that use tempobench logic. Check out the benchmark.py file out on the project github.

from tempobench import Benchmark

bench = Benchmark(
    dataset_path="src/tempobench/data/causal-done.jsonl",
    task="causal",
    model_id="openai/gpt-4o-mini",
    results_dir="benchmark_results",
    console_prints=True,
)

df = bench.evaluate()
print(df.head())

---

Env vars

You will need to have these env vars set for this to work properly.

• OPENROUTER_API_KEY (for --backend openrouter)
• OPENAI_API_KEY (for --backend openai)

---

Results schema (per row)

results_*.jsonl contains:

{
  "model": "openai/gpt-4o-mini",
  "gold": "... (gold JSON) ...",
  "pred": "... (raw text) ...",
  "GT": { "...parsed..." },
  "PRED": { "...parsed..." },
  "correct": true,
  "precision_ap": 1.0,
  "recall_ap": 1.0,
  "F1_ap": 1.0,
  "precision_timestep": 1.0,
  "recall_timestep": 1.0,
  "F1_timestep": 1.0,
  "cost": 0.0023,
  "generation_id": "gen_...",
  "native_prompt_tokens": 123,
  "native_completion_tokens": 45
}

License

MIT (see LICENSE).

If you use this benchmark, please cite our paper!