File size: 8,049 Bytes
a471c50 | 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 | # DeliveryBench evaluation: download to scores
This release adds version-matched targets for all **8,568 requests**, the actual **5,938 training choices**, relative request timestamps, an installable evaluation framework and runnable scoring examples. It uses the existing **722,975 configurations**; existing user/request/configuration identifiers are preserved.
The targets were exported from the source eligibility dataset and checked against the public catalog. Every reference ID exists at the request's location, and every feasible-set count was recomputed. The train/validation/test target counts are **5,938 / 1,239 / 1,391**.
## Install and download
The Dataset is public. No HF token or delivery-platform credential is needed to download or run the literal-parser example.
```sh
python -m pip install huggingface_hub pyarrow
```
```python
from huggingface_hub import snapshot_download
snapshot_download(repo_id="fstxwd/DeliveryBench", repo_type="dataset",
local_dir="DeliveryBench")
```
```sh
python -m pip install DeliveryBench/ada_food_benchmark-0.2.0-py3-none-any.whl
ada-bench validate --data DeliveryBench
```
The source distribution is `evaluation_tools-0.2.0.zip`; its tests, provider configuration, custom-agent example and package metadata are included. `evaluation_requirements.txt` records direct verification dependencies. Use an immutable HF commit/revision when comparing runs.
## Verify the scorer before evaluating a method
```sh
ada-bench score --data DeliveryBench --oracle DeliveryBench --split validation --k 1 --predictions DeliveryBench/evaluation_reference_predictions_validation.jsonl --output runs/reference-check.json
```
This checks the reference predictions against the validation targets and must yield 1,239 scored tasks with validity and selection-success rates of 1.0. **It is a data/scorer integrity check, not an ADA performance result.** Other feasible configurations are also accepted; the reference ID is not the only correct choice.
## Fit the prior and run an agent
```sh
ada-bench fit-prior --data DeliveryBench --output runs/prior.json
ada-bench run --data DeliveryBench --oracle DeliveryBench --prior runs/prior.json --split validation --max-users 1 --output runs/smoke.jsonl
```
The default literal parser needs no model key. The smoke run is one complete user's chronological trajectory, with memory containing only actual earlier accepted choices. Its `.summary.json` describes that subset; it is not a whole-split result.
Remove `--max-users 1` for the complete split. The final evaluation command can use `--split test` after training/configuration selection is frozen. Example full-split run and score:
```sh
ada-bench run --data DeliveryBench --oracle DeliveryBench --prior runs/prior.json --split test --output runs/test.jsonl
ada-bench score --data DeliveryBench --oracle DeliveryBench --split test --predictions runs/test.jsonl --output runs/test-score.json
```
`--agent quality` and `--agent none` are common-interface controls. A third-party agent can implement `reset_user`, `begin_episode`, `act` and `observe_outcome` and use `--agent path/to/agent.py:factory`. See the source archive's `docs/PROTOCOL.md`.
## Optional model-based parsing
The packaged parser accepts an OpenAI-compatible chat endpoint. Set **your own** `ADA_BASE_URL`, `ADA_MODEL` and `ADA_API_KEY`, then add `--parser model` to `ada-bench run`. `.env` is not loaded automatically by the benchmark CLI. These opt-in calls can incur provider costs; no provider secret is in this release. The software-validation runs used the literal parser and made no model calls.
## Submit predictions from any framework
Write one JSON object per request:
```json
{"request_id":"DB_U0001_R001","configuration_ids":["config_..."],"selected_configuration_id":"config_..."}
```
The legacy `episode_id` alias is also accepted. Conflicting aliases, unknown/repeated request IDs and repeated candidate IDs are rejected. The selected ID must appear in the submitted list. Configuration IDs must belong to the request's location. Omitted expected requests remain in the denominator.
`valid_at_k` pads missing candidate slots with zero; `constraint_hit_at_k` measures whether a feasible candidate appears in the first k slots. `selection_success_rate` checks the selected configuration. Static submissions cannot prove user acceptance, latency or model cost. Interactive outputs additionally record environment-verified acceptance and action traces.
## Files
| Resource | Contents |
|---|---|
| `evaluation_labels_train.{parquet,jsonl}` | 5,938 current-episode targets and reference IDs |
| `evaluation_labels_validation.{parquet,jsonl}` | 1,239 targets |
| `evaluation_labels_test.{parquet,jsonl}` | 1,391 openly released targets |
| `evaluation_training_history.{parquet,jsonl}` | Actual source-witness choices for training users only |
| `evaluation_request_context.parquet` | Relative timestamp and sequence for each request |
| `evaluation_manifest.json` | Native-loader file mapping, data hashes, counts and protocol |
| `evaluation_schema.json` | Arrow field definitions |
| `evaluation_validation.json` | Full source-to-public ID and feasibility checks |
| `evaluation_reference_predictions_validation.jsonl` | Reference-only scorer sanity check |
| `ada_food_benchmark-0.2.0-py3-none-any.whl` | Installable framework/scorer |
| `evaluation_tools-0.2.0.zip` | Corresponding source and tests |
| `paper_metric_reanalysis.zip` | Sanitized recorded metrics and no-model point-estimate reanalysis for the available main/crossover/backend studies |
Attributes are `{name,value}` lists. The loader converts them to mappings, maps request fields into the agent protocol and excludes inspection-profile preferences from permitted evaluation observations. All published input/candidate tables retain their current identifiers and contents.
## Timeline, memory and labels
Relative timestamps preserve the source evaluation's whole-week heldout shift (21 days), within-user order/gaps and the recorded tie correction. Training events precede heldout evaluation. Period/hour/weekday remain the published context fields. The frozen catalog does not establish historical availability at these timestamps.
Train/dev goals may support development. Heldout user histories start empty and accumulate actual prior accepted configurations separately for every method/trial. Do not preload the user's inspection preferences, oracle choices or future outcomes as evaluation observations.
**Test labels are intentionally public for open-label research reproducibility.** This is not a hidden-test leaderboard or a claim that the original test users have never been examined. The local harness is a trusted-process SDK, not an isolation boundary for malicious agent code.
## Relation to historical paper scores
The reusable runner uses a deterministic progressive-disclosure user fixture and strict categorical configuration checks. It is a declared benchmark protocol, distinct from historical experiments that used an LLM user, semantic judge and their frozen RRF/baseline implementations. Running the current reference adapter does not regenerate those historical experiment scores.
`paper_metric_reanalysis.zip` separately recomputes recorded point estimates from sanitized attempt-level metrics for the available 1,080 aligned main/ablation runs, 2,400 crossover runs and 144 synthetic-profile backend episodes. It preserves failures and distinguishes 141 validated acceptances from 139 completed acknowledgements. It does not replay model inference, reconstruct missing trajectories, or imply a human-participant study. The archived full-factorial report is descriptive aggregate evidence, not raw trajectory logs.
Offline recommendation/configuration is the measured endpoint. No live order, payment or delivery is performed. Code rights and third-party notices remain those of the accompanying review artifact; this release does not assign a new permissive license to third-party data or code.
|