vi-benchmark-pilot / README.md
initialneil's picture
Dataset card: name the shopping filler sources
5301ab7 verified
|
Raw History Blame Contribute Delete
3.68 kB
---
license: other
pretty_name: VI-Benchmark pilot (video + visual shopping)
extra_gated_prompt: "This dataset contains video derived from third-party research datasets under their own licences (several non-commercial). Access is for non-commercial research only. You agree to follow each source licence listed in video/manifest.jsonl and not to redistribute the footage."
extra_gated_fields:
Name: text
Affiliation: text
Intended use (research, non-commercial): text
I agree to use this dataset for non-commercial research only and to follow the source licences: checkbox
---
# VI-Benchmark pilot
Video-agent benchmark scenarios (VI-ToolSandbox) and agent runs from the pilot. Layout: `video/scenarios/<slug>/`
(task + footage the agent sees), `video/runs/<group>/...` (trajectories, grades), `video/manifest.jsonl` (per scenario:
sources, licence, included or withheld and why). Scenarios whose licence does not allow redistribution, and anything
from Ego4D, are withheld and listed by slug only. Answers/rubrics are not published here.
Revisions are tagged per benchmark version (e.g. `bench-v1.2`).
## Visual shopping: Kimi K3 traces for the Qwen3.5-9B fine-tune
`visual-shopping/` holds the second lane, the **Moving-In Sofa Hunt**: Sam moves from Berkeley to a San Francisco flat
and asks an assistant (with mail, photos, saved memory, old chats and a shop catalog) to pick a sofa plus a rug under
$1,500. The rule that matters is never said: the sofa must fit through a 30 in door. Kimi K3 passes about 80 % of the
frozen set, Qwen3.5-9B 0 %. The traces below are Kimi K3 runs for fine-tuning Qwen3.5-9B on this scenario.
**Pick up in four steps**
1. **Access.** Request access on this repo and on the answer-key repo `CollovLabs/vi-benchmark-pilot-oracle` (both
gated, manual approval). Ask for read access to the GitHub repos `flair-home-stylist/vi-ts-generation` and
`flair-home-stylist/vi-toolsandbox` (runner, agents, judge).
2. **Get the traces.** They are uploaded in batches while the Kimi runs continue (target about 100 passing runs), so
download from `main` and re-run the command to pick up new batches:
```bash
hf download CollovLabs/vi-benchmark-pilot --repo-type dataset --include "visual-shopping/traces/v1/*" --local-dir vi-benchmark-pilot
```
When the runs finish, the final set is tagged `shopping-v1-traces-v1` (use `--revision shopping-v1-traces-v1`).
To pin a batch before that, pass its commit hash as `--revision`.
3. **Train on passing runs.** `traces/v1/runs/<run_id>.jsonl` is one fine-tune example per run: OpenAI-style `tools`
and `messages` (the agent's view: system, Sam's turns, assistant tool calls, tool results; Kimi's reasoning in a
separate `reasoning` field; images by path into `traces/v1/images/`). `passing.txt` lists the runs to train on;
failing runs are kept for analysis. `manifest.jsonl` has one row per run (variant, paraphrase seed, pass, per-item
verdicts, tokens).
4. **Evaluate.** The variant `moving-in-s-rich-a7h` is **held out**: it has no traces, never train on it. To run Qwen
before and after fine-tuning, download the world and scenarios (tag `shopping-v1`) plus the answer key, assemble a
scenario root and run the runner; the exact commands, the grading (code checks + an evidence-grounded rubric judge,
pass = all of both) and the frozen numbers are in `visual-shopping/README.md` (sections Reproduce, Train / test
split, Traces).
Licences for the shopping world (synthetic files, AppWorld mail as filler, HierText and WorldVQA filler photos; some carry redistribution caveats) are
listed in `visual-shopping/README.md` under Licences and sources.