Text Classification
Transformers
Safetensors
qwen3_5
image-text-to-text
typed-decisions
calibrated-classification
system-one
classification
structured-prediction
candidate-logit
jev
single-forward-pass
multilingual
commercial-use
Instructions to use wayfind/metask-jev-4b-policy-mix with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use wayfind/metask-jev-4b-policy-mix with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-classification", model="wayfind/metask-jev-4b-policy-mix")# Load model directly from transformers import AutoProcessor, AutoModelForMultimodalLM processor = AutoProcessor.from_pretrained("wayfind/metask-jev-4b-policy-mix") model = AutoModelForMultimodalLM.from_pretrained("wayfind/metask-jev-4b-policy-mix", device_map="auto") - Notebooks
- Google Colab
- Kaggle
Download selftest.sh from wayfind/metask-jev-4b-policy-mix: direct link, hf CLI and curl.
- Browser
- Download file 3.33 kB
-
https://huggingface.co/wayfind/metask-jev-4b-policy-mix/resolve/main/selftest.sh
- Command line
-
hf download hf://wayfind/metask-jev-4b-policy-mix/selftest.sh
-
curl -L -o selftest.sh https://huggingface.co/wayfind/metask-jev-4b-policy-mix/resolve/main/selftest.sh
3.33 kB
| # metask-jev-4b self-test: reproduce the published JevBench numbers on your machine. | |
| # Device auto-detected (CUDA / Apple Silicon MPS / CPU). Resumable. | |
| # Results: ~/metask-jev/bench_results/*.jsonl (per-item) + summary table below. | |
| set -uo pipefail | |
| GREEN='\033[1;32m'; RED='\033[1;31m'; NC='\033[0m' | |
| say() { printf "\n${GREEN}[metask-jev]${NC} %s\n" "$*"; } | |
| err() { printf "\n${RED}[metask-jev]${NC} %s\n" "$*"; } | |
| DIR="$HOME/metask-jev" | |
| RESULTS="$DIR/bench_results" | |
| mkdir -p "$RESULTS" | |
| # ---------- environment ---------- | |
| if [ ! -d "$DIR/.venv" ]; then | |
| say "no install detected β run the one-command install first:" | |
| echo " curl -fsSL https://raw.githubusercontent.com/metask-ai/metask-jev/main/install.sh | bash" | |
| exit 1 | |
| fi | |
| source "$DIR/.venv/bin/activate" | |
| if ! python -c "from transformers import Qwen3_5ForConditionalGeneration" 2>/dev/null; then | |
| err "transformers is missing the Qwen3_5 class β upgrade: pip install -U transformers" | |
| exit 1 | |
| fi | |
| # ---------- device ---------- | |
| if command -v nvidia-smi >/dev/null 2>&1; then DEV="cuda" | |
| elif [ "$(uname)" = "Darwin" ]; then DEV="mps" | |
| else DEV="cpu"; fi | |
| say "device: $DEV" | |
| # ---------- task files ---------- | |
| cd "$DIR" | |
| for f in easy original hard; do | |
| if [ ! -s "$f.jsonl" ]; then | |
| say "fetching JevBench $f.jsonl..." | |
| curl -fsSL "https://raw.githubusercontent.com/fstandhartinger/jevbench/main/datasets/public/$f.jsonl" -o "$f.jsonl" | |
| fi | |
| done | |
| # ---------- benchmark (3 tiers, resumable) ---------- | |
| if [ ! -f "run_bench_local.py" ]; then | |
| say "fetching the benchmark runner..." | |
| curl -fsSL "https://raw.githubusercontent.com/metask-ai/metask-jev/main/run_bench_local.py" -o "run_bench_local.py" | |
| fi | |
| say "running: easy 48 + judge 72 + hard 111 (resumable; ~25-40 min on MPS)" | |
| python run_bench_local.py --tiers "easy judge hard" \ | |
| --results-dir "$RESULTS" \ | |
| --model wayfind/metask-jev-4b-policy-mix 2>&1 | grep -vE "Loading|warn|falling back" | |
| # ---------- summary ---------- | |
| echo | |
| say "βββββββββββ results βββββββββββ" | |
| printf "| %-8s | %-12s | %-9s |\n" "tier" "correct/att" "accuracy" | |
| printf "|----------|--------------|----------|\n" | |
| GC=0; GN=0 | |
| # tier name -> JevBench task file name (the judge tier lives in original.jsonl) | |
| tier_file() { case "$1" in easy) echo easy;; judge) echo original;; hard) echo hard;; esac; } | |
| for pair in "easy:48" "judge:72" "hard:111"; do | |
| T=${pair%%:*}; N=${pair##*:} | |
| F=$(tier_file "$T") | |
| OK=0; C=0 | |
| if [ -s "$RESULTS/$F.jsonl" ]; then | |
| while IFS= read -r line; do | |
| R=$(echo "$line" | python3 -c "import json,sys; r=json.load(sys.stdin); print(r['ok'], str(r['prediction'])==str(r['expected']))") | |
| [ "$(echo $R | cut -d' ' -f1)" = "True" ] && OK=$((OK+1)) && { [ "$(echo $R | cut -d' ' -f2)" = "True" ] && C=$((C+1)); } | |
| done < "$RESULTS/$F.jsonl" | |
| fi | |
| GN=$((GN+OK)); GC=$((GC+C)) | |
| if [ "$OK" -gt 0 ]; then | |
| ACC=$(python3 -c "print(f'{$C/$OK*100:.1f}%')") | |
| printf "| %-8s | %-12s | %-9s |\n" "$T" "$C/$OK" "$ACC" | |
| else | |
| printf "| %-8s | %-12s | %-9s |\n" "$T" "skipped" "-" | |
| fi | |
| done | |
| printf "| %-8s | %-12s | %-9s |\n" "total" "$GC/$GN" "$(python3 -c "print(f'{$GC/$GN*100:.1f}%' if $GN else '-')")" | |
| echo | |
| say "published reference: 80.1% @4096 ctx (see the model card for the full tables)" | |
| say "per-item results: $RESULTS/*.jsonl" |