davis901's picture
AndroidFlux evaluation outputs
02039a4
|
Raw History Blame Contribute Delete
4.15 kB
# Run manifests
Input manifests for the replay / failure-recovery evaluation cells published in this
repository. Each cell under `llm_verified/`, `human_verified/`, and `supplementary/` was produced
by running one of these manifests under one condition.
## Which manifest goes with which cell
| Axis (cell path prefix) | Manifest | Entries |
|---|---|---|
| `llm_verified/replay_312/{clean,t,t_minus_1}` | `lv312_manifest.json` | 312 |
| `llm_verified/replay_312/t_plus_1_pos` | `lv220_tplus1_positive.json` | 220 |
| `llm_verified/replay_312/t_plus_1_neg` | `lv220_tplus1_negative.json` | 220 |
`lv312_manifest.shard_0N_of_04.json` are the canonical 4-way split of `lv312_manifest.json`
(78 entries each). The split is fixed: an entry's shard is a property of the manifest, not of
the run. **Re-running a cell with a different split places results in different shard
directories; the runner only skips episodes already present in its own shard directory, so a
changed split silently re-runs work and produces duplicate entries.** Use the same split on
every run of a cell.
## Denominator and scoring
`success_rate = success_count / expected_entries`, where `expected_entries` is the manifest
entry count for that axis (312 or 220) — not the number of result files found. A cell whose
`validation.json` reports a non-zero `unexpected_result` count was validated against a
different entry set than the one it claims; its numbers are not comparable to other cells.
## Arm convention for t+1
The positive and negative arms share the same 220 tasks and are distinguished **only by the
`output_key` suffix** (`__tplus1_positive` / `__tplus1_negative`). Neither the directory name
nor the `condition` field carries the arm, so joining two t+1 cells on `(model, task)` alone
cannot separate them.
## Join key
An episode's identity is the **directory-name prefix** (the model key, up to the first `__`)
plus the `task_template` field **inside** `episode_result.json`. Directory names are
truncated goals for some generations and canonical task names for others, so the directory
name alone is not a key.
## Conditions
| Condition | `--start-mode` | `--replay-mode` | History arguments |
|---|---|---|---|
| `clean` | `clean_start` | `through_selected_recovery_error_step` | — |
| `t` (`error_t`) | `replay` | `through_selected_recovery_error_step` | `--history-prime-from-replay --use-prebuilt-replay-memory --replay-memory-metadata-root <root> --no-inject-replay-context` |
| `t_minus_1` | `replay` | `through_selected_recovery_last_correct_step` | same as `t` |
| `t_plus_1` | `replay` | `through_selected_recovery_next_action` | `--history-prime-from-replay --no-inject-replay-context` |
**`t_plus_1` does not take prebuilt replay memory.** Passing
`--use-prebuilt-replay-memory` on the t+1 conditions yields episodes with zero recovery steps.
Budget arguments are common to all conditions: `--dynamic-replay-recovery-budget`,
`--dynamic-replay-budget-reference entry_t_replayed`, `--history-budget-threshold-ratio 0.5`,
`--short-history-budget-multiplier 1.5`, `--long-history-budget-multiplier 2.0`,
`--no-subtract-replay-actions-from-step-budget`.
## Placeholders
`<REPO_ROOT>` and `<MODEL_ROOT>` stand for the checkout and model directories on the machine
that produced a run. The fields containing them (`materialized_original_task_dir` and
`raw_task_record.task_dir`) are provenance records: the evaluation runner does not read them.
Every field the runner does read — `task_dir`, `task_dir_relative`, `model_name`,
`canonical_task_name`, `task_name`, `output_key`, `replay_shard_assignment` — is unchanged.
## Model identity
Serve each model under the id the cell was produced with, and request that exact string.
A mismatched model name returns 404, which the agent swallows: the episode is written with an
empty response and the cell completes at full count with no error recorded. The empty
response field is the only signal.
`ma3_7b` and `gui_owl_7b` load the **same weights**; they differ only in the agent
(`mobile_agent_v3` vs `gui_owl_vllm`). Matching the served model id alone does not distinguish
the two cells.