Scaled evals

benchmarks/ppo/accelerate_configs/deepspeed_zero2.yaml

@@ -12,10 +12,11 @@ machine_rank: 0
 main_training_function: main
 mixed_precision: 'bf16'
 num_machines: 1
-num_processes: 1
+num_processes: 2
 rdzv_backend: static
 same_network: true
 tpu_env: []
 tpu_use_cluster: false
 tpu_use_sudo: false
 use_cpu: false
+main_process_port: 0

jobs/validate_all_static.sh

@@ -2,6 +2,7 @@
 #SBATCH --job-name=validate_static_all
 #SBATCH --partition=unkillable-cpu
 #SBATCH --cpus-per-task=2
+#SBATCH --mem=16G
 #SBATCH --time=24:00:00
 #SBATCH --output=slurm-%j.out
 #SBATCH --error=slurm-%j.err
@@ -10,130 +11,24 @@
 
 source .env
 
-uv run aif \
-validate \
---max_concurrency 256 \
-data/70B_15_generation/education_qna_direct/*/data.json \
-data/70B_15_validation/70B/education_qna_direct/validate.json \
---no-validate-diversity \
---model Meta-Llama-3.1-70B-Instruct \
-&& \
-uv run aif \
-validate \
---max_concurrency 256 \
-data/70B_15_generation/education_qna_eli5/*/data.json \
-data/70B_15_validation/70B/education_qna_eli5/validate.json \
---no-validate-diversity \
---model Meta-Llama-3.1-70B-Instruct \
-&& \
-uv run aif \
-validate \
---max_concurrency 256 \
-data/70B_15_generation/education_qna_expert/*/data.json \
-data/70B_15_validation/70B/education_qna_expert/validate.json \
---no-validate-diversity \
---model Meta-Llama-3.1-70B-Instruct \
-&& \
-uv run aif \
-validate \
---max_concurrency 256 \
-data/70B_15_generation/education_qna_hinted/*/data.json \
-data/70B_15_validation/70B/education_qna_hinted/validate.json \
---no-validate-diversity \
---model Meta-Llama-3.1-70B-Instruct \
-&& \
-uv run aif \
-validate \
---max_concurrency 256 \
-data/70B_15_generation/education_summary_eli5/*/data.json \
-data/70B_15_validation/70B/education_summary_eli5/validate.json \
---no-validate-diversity \
---model Meta-Llama-3.1-70B-Instruct \
-&& \
-uv run aif \
-validate \
---max_concurrency 256 \
-data/70B_15_generation/education_summary_expert/*/data.json \
-data/70B_15_validation/70B/education_summary_expert/validate.json \
---no-validate-diversity \
---model Meta-Llama-3.1-70B-Instruct \
-&& \
-uv run aif \
-validate \
---max_concurrency 256 \
-data/70B_15_generation/politics_generate_long/*/data.json \
-data/70B_15_validation/70B/politics_generate_long/validate.json \
---no-validate-diversity \
---model Meta-Llama-3.1-70B-Instruct \
-&& \
-uv run aif \
-validate \
---max_concurrency 256 \
-data/70B_15_generation/politics_generate_short/*/data.json \
-data/70B_15_validation/70B/politics_generate_short/validate.json \
---no-validate-diversity \
---model Meta-Llama-3.1-70B-Instruct \
-&& \
-uv run aif \
-validate \
---max_concurrency 256 \
-data/70B_15_generation/politics_qna_eli5/*/data.json \
-data/70B_15_validation/70B/politics_qna_eli5/validate.json \
---no-validate-diversity \
---model Meta-Llama-3.1-70B-Instruct \
-&& \
-uv run aif \
-validate \
---max_concurrency 256 \
-data/70B_15_generation/politics_qna_expert/*/data.json \
-data/70B_15_validation/70B/politics_qna_expert/validate.json \
---no-validate-diversity \
---model Meta-Llama-3.1-70B-Instruct \
-&& \
-uv run aif \
-validate \
---max_concurrency 256 \
-data/70B_15_generation/politics_summary_eli5/*/data.json \
-data/70B_15_validation/70B/politics_summary_eli5/validate.json \
---no-validate-diversity \
---model Meta-Llama-3.1-70B-Instruct \
-&& \
-uv run aif \
-validate \
---max_concurrency 256 \
-data/70B_15_generation/politics_summary_expert/*/data.json \
-data/70B_15_validation/70B/politics_summary_expert/validate.json \
---no-validate-diversity \
---model Meta-Llama-3.1-70B-Instruct \
-&& \
-uv run aif \
-validate \
---max_concurrency 256 \
-data/70B_15_generation/politics_generate_short/*/data.json \
-data/70B_15_validation/70B/politics_generate_short/validate.json \
---no-validate-diversity \
---model Meta-Llama-3.1-70B-Instruct \
-&& \
-uv run aif \
-validate \
---max_concurrency 256 \
-data/70B_15_generation/politics_qna_eli5/*/data.json \
-data/70B_15_validation/70B/politics_qna_eli5/validate.json \
---no-validate-diversity \
---model Meta-Llama-3.1-70B-Instruct \
-&& \
-uv run aif \
-validate \
---max_concurrency 256 \
-data/70B_15_generation/politics_summary_eli5/*/data.json \
-data/70B_15_validation/70B/politics_summary_eli5/validate.json \
---no-validate-diversity \
---model Meta-Llama-3.1-70B-Instruct \
-&& \
-uv run aif \
-validate \
---max_concurrency 256 \
-data/70B_15_generation/tech_healthcare_summary_expert/*/data.json \
-data/70B_15_validation/70B/tech_healthcare_summary_expert/validate.json \
---no-validate-diversity \
---model Meta-Llama-3.1-70B-Instruct
+MODEL="gpt-4o-mini"
+MAX_CONC=256
+
+# list all sub‐tasks
+tasks=(
+  ultra-hh-sampled
+)
+
+for t in "${tasks[@]}"; do
+  echo "Validating $t..."
+  uv run aif validate \
+    --max_concurrency "$MAX_CONC" \
+    "data/$t.json" \
+    "data/$t-validate-no-diversity.json" \
+    --no-validate-diversity \
+    --no-validate-embedding-diversity \
+    --model "$MODEL" \
+    || { echo "Validation failed on $t"; exit 1; }
+done
+
+echo "All validations completed successfully."

jobs/validate_all_static_diversity.sh

@@ -0,0 +1,71 @@
+#!/bin/bash
+#SBATCH --job-name=validate_static_all_diversity_external
+#SBATCH --partition=main
+#SBATCH --gres=gpu:a100l:1
+#SBATCH --mem=36G
+#SBATCH --cpus-per-task=6
+#SBATCH --time=24:00:00
+#SBATCH --output=slurm-%j.out
+#SBATCH --error=slurm-%j.err
+#SBATCH --mail-type=ALL
+#SBATCH --mail-user=
+
+cd
+module load python/3.10
+module load cuda/12.6.0
+source .venv/bin/activate
+echo "Starting vLLM server..."
+uv run vllm serve BAAI/bge-m3 --dtype float16 --api-key openai --task embed &
+
+# Save server process ID
+SERVER_PID=$!
+
+echo "Waiting for server to start..."
+while true; do
+  echo "Checking if server is up..."
+  RESPONSE=$(curl -s http://localhost:8000/v1/models -H "Authorization: Bearer openai" 2>&1)
+
+  if [[ "$RESPONSE" == *"data"* ]]; then
+    echo "Server is up and running!"
+    break
+  fi
+
+  # Check if server is still running
+  if ! kill -0 $SERVER_PID 2>/dev/null; then
+    echo "Server process died unexpectedly"
+    exit 1
+  fi
+
+  echo "Server not ready yet. Waiting 5 seconds..."
+  sleep 5
+done
+
+deactivate
+cd projects/AIF-Gen
+source .env
+
+echo "Starting validation process..."
+
+# list all sub‐tasks
+tasks=(
+  cppo-reward-sampled
+  cppo-rl-sampled
+  ultra-hh-sampled
+)
+
+for t in "${tasks[@]}"; do
+  echo "Validating $t..."
+  uv run aif validate \
+    "data/$t.json" \
+    "data/$t-validate-diversity.json" \
+    --no-validate-diversity \
+    --no-validate-count \
+    --no-validate-entropy \
+    --no-validate-llm-judge \
+    --embedding-model "BAAI/bge-m3" \
+    --embedding-batch-size 256 \
+    --max_concurrency 16 \
+    || { echo "Validation failed on $t"; exit 1; }
+done
+
+echo "All validations completed successfully."

pyproject.toml

@@ -26,6 +26,7 @@ dependencies = [
     "pydantic>=2.10.4",
     "pytest-asyncio>=0.25.3",
     "pytest-mock>=3.14.0",
+    "setuptools>=75.8.2",
     "torch==2.3.0",
     "types-pyyaml>=6.0.12.20241230",
 ]
@@ -40,6 +41,8 @@ dev = [
     "ruff>=0.7.3",
     "sphinx>=7.4.7",
     "sphinx-rtd-theme>=3.0.2",
+    "ipykernel>=6.29.5",
+    "matplotlib>=3.10.1",
 ]
 benchmarks-dpo = [
     "datasets>=3.2.0",