File size: 8,435 Bytes
26e48a5
 
 
 
 
 
 
 
 
 
f138571
 
 
 
 
 
 
26e48a5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fa3f6f8
26e48a5
fa3f6f8
26e48a5
 
 
 
 
fa3f6f8
26e48a5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fa3f6f8
26e48a5
 
 
 
 
 
 
 
 
 
 
fa3f6f8
26e48a5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fa3f6f8
 
26e48a5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
# BenchCheck-Pool: full-item-pool screening, steps 1-3 (run package)

This package lets an agent on a separate GPU machine run the three model-based screening steps of the
BenchCheck full-item pool and send the per-item results back. Everything needed is here except the
two open-weight models (downloaded from the Hugging Face hub) and the frame cache, which sits in the
companion dataset `GMLRVigil/BenchCheck-Pool-Frames` (public).

Produce: `results/pool/step1/*.jsonl`, `results/pool/step2/*.jsonl`, `results/pool/step3/*.jsonl`
(one row per item, append-only, resumable), packed and uploaded as described in section 8.

**Fix 2026-09-11:** the first upload of `supplement_20260910/code/` (and of the main package's `code/`) lacked
`code/exp/pipeline/` (chance.py, scoring.py, stage_d_blind.py). `pool_steps.eligible` imports
`chance._yes_no_options` (the yes/no-option rule of 2026-09-10) from that folder, so `--plan-only` and the run stop with
`ModuleNotFoundError: No module named 'chance'`. Both repos now carry the folder; re-download
`supplement_20260910/code/` and copy it over `pool/code/` (the three files are verbatim copies of our pipeline, the
screening rule is unchanged). Nothing else in the package changed.

## 1. What the three steps do (do not change any of it)

Items: multiple-choice questions of 139 video benchmarks (299,366 after the step-0 dedup, table in
`items/`). Each step asks a Qwen3-VL model the question and reads the log-probability of every option
letter as the first generated token (assistant turn prefilled with `Answer:`, `max_tokens=1`,
`logprobs=true`, `top_logprobs=20`). An item leaves the pool when the model is right with a margin
above log 2:

| step | model | input | removal rule |
|---|---|---|---|
| 1 blind | Qwen3-VL-8B-Instruct | question + options, no frames, 4 option permutations per item | ≥ 3 of 4 permutations pick the gold letter and mean margin > log 2 |
| 2 single frame | Qwen3-VL-8B-Instruct | the middle frame (448 px long side) + question | gold letter is argmax and margin > log 2 |
| 3 32 frames | Qwen3-VL-2B-Instruct | 32 uniform frames (448 px) + question | same |

Step 2 runs only on items that survived step 1 and have frames; step 3 only on survivors of step 2.
Prompts, permutation seeds, frame selection, resolution, model revisions and the margin rule are
fixed in the code (`code/exp/analysis/pool/pool_steps.py`, `pool_common.py`, `pool_prompts.py`); the
paper protocol depends on them, so run the code as is. `code/exp/analysis/pool/TASK.md` is the
original task note (Chinese).

## 2. Hardware and software

Target: 4x NVIDIA B200 (180 GB each); also runs on 32 GB GPUs. Per GPU one vLLM server and one client shard.
VRAM: Qwen3-VL-8B bf16 weights ~16.4 GB, Qwen3-VL-2B ~4.4 GB; the launcher uses
`--gpu-memory-utilization 0.85`. On 32 GB GPUs, if vLLM fails with out-of-memory at start-up, lower `MAX_SEQS`
in `code/run_remote.sh` (e.g. 64 -> 32 for step 1); do not lower the resolution or the frame count.

```bash
# Python 3.12 environment
python3 -m venv pool-env && source pool-env/bin/activate
pip install -r code/requirements.txt          # vllm 0.11.0 + torch 2.8 (CUDA 12.8 wheels; Blackwell B200 / RTX 5090)
pip install -U "huggingface_hub[cli]"
# models (pinned revisions; served under their hub names)
huggingface-cli download Qwen/Qwen3-VL-8B-Instruct --revision 0c351dd01ed87e9c1b53cbc748cba10e6187ff3b
huggingface-cli download Qwen/Qwen3-VL-2B-Instruct --revision 89644892e4d85e24eaac8bacfd4f463576704203
```

Disk: items 80 MB, frames ~35 GB unpacked (+35 GB for the tar shards until you delete them),
results < 2 GB, model weights ~21 GB.

## 3. Get the data

```bash
huggingface-cli download GMLRVigil/BenchCheck-Pool --repo-type dataset --local-dir pool
cd pool
mkdir -p results/pool && cp items/*.parquet items/*.csv items/*.jsonl results/pool/
# frames (public repo)
huggingface-cli download GMLRVigil/BenchCheck-Pool-Frames --repo-type dataset --local-dir frames_tar
python3 - <<'PY'
import json, hashlib, sys
m = json.load(open("frames_tar/frames_manifest.json"))
for s in m["shards"]:
    h = hashlib.sha256(open("frames_tar/" + s["file"], "rb").read()).hexdigest()
    print(s["file"], "OK" if h == s["sha256"] else "SHA256 MISMATCH"); assert h == s["sha256"]
PY
mkdir -p frames/pool_frames && for t in frames_tar/frames_*.tar; do tar -xf "$t" -C frames/pool_frames; done
ls frames/pool_frames | wc -l      # expect the n_videos value of frames_tar/frames_manifest.json (67,021)
```

After this the package root `pool/` holds `code/`, `items/`, `results/pool/`, `frames/pool_frames/`.

## 4. Run

`code/run_remote.sh` starts one vLLM server per GPU (ports 8001+i), waits until they answer, runs one
client shard per GPU, then prints the remaining work. Run the steps in order; each command can be
re-run at any time and continues where it stopped (rows are keyed by item_id).

```bash
cd pool
STEP=1 bash code/run_remote.sh --plan-only    # 299,366 items expected before the first run
STEP=1 bash code/run_remote.sh                # ~1-2 h on 4x B200 (estimate)
STEP=2 bash code/run_remote.sh --plan-only    # survivors of step 1 that have frames (upper bound 183,196)
STEP=2 bash code/run_remote.sh                # ~1.5-2.5 h (estimate)
STEP=3 bash code/run_remote.sh --plan-only
STEP=3 bash code/run_remote.sh                # ~2-4 h (estimate); 2B model, 32 images per request
```

Environment variables the launcher accepts: `NGPU` (default: all visible GPUs), `WORKERS` (client
threads per GPU; defaults 32 / 16 / 8 for steps 1 / 2 / 3), `PORT0`, `GPU_UTIL`, `PY`, `VLLM`.
Logs: `logs/vllm_s<STEP>_gpu<i>.log`, `logs/client_s<STEP>_gpu<i>.log`.

The time estimates come from throughput on the origin cluster's smaller GPU slices and were not
measured on this hardware; step 1 calibrates them (its client log prints items/s).

## 5. Progress and expected counts

`STEP=N bash code/run_remote.sh --plan-only` prints `REMAINING_TOTAL=<n>`; a step is complete when it
prints 0. Step 1 covers 299,366 items = 1,197,464 forwards. Of these items, 183,196 (67,021 videos)
have frames in this snapshot; the other 116,170 items reference videos that were not yet downloaded on
the origin cluster (video_id starting with `k:`), so steps 2 and 3 skip them here. They will be finished
on the origin cluster; nothing to do about them on your side.

Sanity checks:
- Step 1 removes a minority of items per benchmark (blind removal on the 300-item samples was mostly
  20-60 %); a benchmark with ~100 % removal or ~0 % removal deserves a look at its client log.
- Every step-1 item has 4 permutation forwards (`n_perm` / per-permutation fields in the row).
- Rows with `status != "ok"` are retried automatically up to the limit in the code; a few hundred
  permanent errors in a million forwards is normal, thousands is not (check the vLLM log for OOM).

## 6. Files that must not be edited

`code/exp/analysis/pool/*.py` (protocol), `items/*` (inputs). The launcher `code/run_remote.sh` may be
adapted to the machine (ports, GPU count, paths) but not the vLLM model arguments other than
`--max-num-seqs` and `--gpu-memory-utilization`.

## 7. Known limits

- vLLM 0.11.0 with the CUDA 12.8 torch 2.8 wheels (Blackwell B200 / RTX 5090). If `pip install vllm==0.11.0`
  pulls a torch without Blackwell support, install torch 2.8.0 from the cu128 index first.
- `--max-logprobs 20` and `continue_final_message` are vLLM features; the client will not work against
  other serving stacks.
- The frame cache is a snapshot; missing directories for some `video_id` values are expected (see 5).

## 8. Return the results

```bash
cd pool
tar -czf pool_returns_$(hostname)_$(date +%Y%m%d).tar.gz results/pool/step1 results/pool/step2 results/pool/step3 logs
sha256sum pool_returns_*.tar.gz > pool_returns.sha256
# with a write token for GMLRVigil:
huggingface-cli upload GMLRVigil/BenchCheck-Pool pool_returns_$(hostname)_$(date +%Y%m%d).tar.gz returns/pool_returns_$(hostname)_$(date +%Y%m%d).tar.gz --repo-type dataset
huggingface-cli upload GMLRVigil/BenchCheck-Pool pool_returns.sha256 returns/pool_returns_$(hostname)_$(date +%Y%m%d).sha256 --repo-type dataset
```

If you have no write token, send the tarball and its sha256 by any file transfer. In the report,
state per step: items done, items removed, rows with status error (count and example item_ids),
wall time and items/s per GPU, and anything you changed in `run_remote.sh`.