Stable, product-facing helper scripts. Research-only tooling lives under
../research/scripts/.
| Script | Purpose |
|---|---|
bootstrap_gpt2_demo.py |
Download/convert default GPT-2 demo weights + tokenizer (./espresso / prepare) |
export_gpt2_coreml.py |
Export GPT-2 Core ML trunk baselines used by compare / bench |
run_gpt2_coreml_reference.py |
Optional Python Core ML reference runner (Swift native fallback exists) |
ensure_coreml_model.sh |
Ensure a CoreML baseline model is available for compare benches |
generate_coreml_model.py |
Build CoreML packages used in comparisons |
reproduce_local_real_artifact_claim.sh |
Reproduce the public real-artifact claim path |
run_power_benchmark.sh |
Power / energy-oriented local runs |
generate-benchmark-dashboard.sh |
Regenerate docs/benchmarks.md from latest.json |
assert_readme_claims.py |
CI guard: README numbers must match latest.json |
assert_demo_helpers.py |
CI guard: GPT-2 demo helper scripts must remain tracked |
convert_weights_gpt2.py / convert_weights_llama.py |
Weight conversion helpers |
convert_qwen25_05b_to_esp.py |
Convert Qwen2.5-0.5B-Instruct (default) or Qwen2.5-1.5B-Instruct (--model) to a packed .esp bundle |
qwen25_pytorch_reference.py |
PyTorch fp32 oracle: layer-parity, logit-parity (probe + NumPy LM head), greedy fixtures |
chat_plain_name_recall.py |
10-turn ./espresso chat --plain --greedy smoke: turn 1 names Ada, a later turn must recall it. Writes docs/qwen15b-chat-name-recall.txt. |
mlx_qwen_stream.py |
Persistent MLX JSON-line driver for ./espresso chat --vs mlx. Native fp16/bf16; quantized loads require --allow-quant. |
chat_vs_mlx_four_turn.py |
Scripted 4-turn ./espresso chat --vs mlx --plain --greedy --power driver. Prints the post-run scoreboard; does not record a GIF. |
espresso_llama_weights.py / stories_model_identity.py |
Stories/Llama weight identity helpers |
export_llama_coreml.py |
Export Llama-family CoreML baselines |
benchmark-prompts.txt / stories_prompt_suite.txt |
Fixed prompt suites |
Python unit tests for product helpers: scripts/tests/.
Autoresearch suite runners, Stories distillation, factored-head packaging, and
rejected draft packaging live in research/scripts/. Do not link them from the
README product journey.