synth-ai 0.2.8.dev4__py3-none-any.whl → 0.2.23.dev3__py3-none-any.whl

examples/blog_posts/mipro/run_mipro_banking77_pipeline_gpt41mini.sh

@@ -0,0 +1,173 @@
+#!/bin/bash
+# Run MIPROv2 optimisation for the multi-step Banking77 pipeline using gpt-4.1-mini as policy model
+
+set -euo pipefail
+
+echo "🔬 Running MIPROv2 on Banking77 Pipeline (Policy: gpt-4.1-mini)"
+echo "================================================================="
+
+REPO_ROOT="$(cd "$(dirname "$0")/../../.." && pwd)"
+cd "$REPO_ROOT"
+
+SAVED_BACKEND_BASE_URL="${BACKEND_BASE_URL:-}"
+SAVED_SYNTH_BASE_URL="${SYNTH_BASE_URL:-}"
+
+_load_env_file() {
+    local env_file="$1"
+    if [ -f "$env_file" ]; then
+        echo "📝 Loading environment variables from $env_file..."
+        while IFS= read -r line || [ -n "$line" ]; do
+            # Skip comments and empty lines
+            [[ "$line" =~ ^[[:space:]]*# ]] && continue
+            [[ -z "${line// }" ]] && continue
+            # Only process lines with '=' character
+            [[ ! "$line" =~ = ]] && continue
+            
+            if [[ "$line" =~ ^[[:space:]]*BACKEND_BASE_URL= ]]; then
+                if [ -z "$SAVED_BACKEND_BASE_URL" ]; then
+                    export "$line" 2>/dev/null || true
+                else
+                    echo "   ⚠️  Skipping BACKEND_BASE_URL from .env (using environment value)"
+                fi
+            elif [[ "$line" =~ ^[[:space:]]*SYNTH_BASE_URL= ]]; then
+                if [ -z "$SAVED_SYNTH_BASE_URL" ]; then
+                    export "$line" 2>/dev/null || true
+                else
+                    echo "   ⚠️  Skipping SYNTH_BASE_URL from .env (using environment value)"
+                fi
+            elif [[ "$line" =~ ^[[:space:]]*[A-Za-z_][A-Za-z0-9_]*= ]]; then
+                export "$line" 2>/dev/null || true
+            fi
+        done < "$env_file"
+    fi
+}
+
+_load_env_file "$REPO_ROOT/.env"
+_load_env_file "$REPO_ROOT/examples/rl/.env"
+
+if [ -n "$SAVED_BACKEND_BASE_URL" ]; then
+    export BACKEND_BASE_URL="$SAVED_BACKEND_BASE_URL"
+    echo "✅ Using BACKEND_BASE_URL from environment: $BACKEND_BASE_URL"
+fi
+if [ -n "$SAVED_SYNTH_BASE_URL" ]; then
+    export SYNTH_BASE_URL="$SAVED_SYNTH_BASE_URL"
+    echo "✅ Using SYNTH_BASE_URL from environment: $SYNTH_BASE_URL"
+fi
+
+if [ -z "${SYNTH_API_KEY:-}" ]; then
+    echo "❌ Error: SYNTH_API_KEY not set"
+    exit 1
+fi
+
+if [ -z "${ENVIRONMENT_API_KEY:-}" ]; then
+    echo "❌ Error: ENVIRONMENT_API_KEY not set"
+    exit 1
+fi
+
+if [ -z "${GROQ_API_KEY:-}" ] && [ -z "${OPENAI_API_KEY:-}" ]; then
+    echo "⚠️  Warning: Neither GROQ_API_KEY nor OPENAI_API_KEY is set."
+    echo "The policy defaults to Groq-hosted OSS models. Set GROQ_API_KEY to avoid failures."
+fi
+
+if [ -z "${OPENAI_API_KEY:-}" ]; then
+    echo "⚠️  Warning: OPENAI_API_KEY not set (required for meta-model gpt-4.1-mini)."
+    read -p "Continue anyway? (y/N) " -n 1 -r
+    echo
+    if [[ ! $REPLY =~ ^[Yy]$ ]]; then
+        exit 1
+    fi
+fi
+
+if [ -z "${BACKEND_BASE_URL:-}" ]; then
+    BACKEND_BASE_URL="http://localhost:8000"
+fi
+BACKEND_URL="$BACKEND_BASE_URL"
+BACKEND_URL_NO_API="${BACKEND_URL%/api}"
+
+echo ""
+echo "🔧 Debug Info:"
+echo "   BACKEND_BASE_URL current: $BACKEND_BASE_URL"
+echo "   BACKEND_URL: $BACKEND_URL"
+echo ""
+
+echo "✅ SYNTH_API_KEY: ${SYNTH_API_KEY:0:20}..."
+echo "✅ ENVIRONMENT_API_KEY: ${ENVIRONMENT_API_KEY:0:20}..."
+if [ -n "${GROQ_API_KEY:-}" ]; then
+    echo "✅ GROQ_API_KEY: ${GROQ_API_KEY:0:20}..."
+fi
+if [ -n "${OPENAI_API_KEY:-}" ]; then
+    echo "✅ OPENAI_API_KEY: ${OPENAI_API_KEY:0:20}..."
+fi
+echo "✅ Backend URL: $BACKEND_URL"
+echo ""
+
+CONFIG_PATH="examples/blog_posts/mipro/configs/banking77_pipeline_mipro_gpt41mini_local.toml"
+
+# ALWAYS read task_app_url from TOML (never use any pre-existing shell variable)
+# Override only if TASK_APP_URL_OVERRIDE is explicitly set
+if [ -n "${TASK_APP_URL_OVERRIDE:-}" ]; then
+    TASK_APP_URL="$TASK_APP_URL_OVERRIDE"
+    echo "📝 Using OVERRIDE task app URL: $TASK_APP_URL"
+else
+    # Extract task_app_url from TOML using grep/sed (no Python dependencies)
+    # Pattern: task_app_url = "https://..."
+    TASK_APP_URL="$(grep "^task_app_url" "$CONFIG_PATH" | sed -E 's/^[^=]*=[[:space:]]*"([^"]*)".*/\1/' | head -1)"
+    if [ -z "$TASK_APP_URL" ]; then
+        echo "❌ ERROR: task_app_url not found in $CONFIG_PATH" >&2
+        echo "   Please ensure the config file contains: task_app_url = \"...\"" >&2
+        exit 1
+    fi
+    echo "📝 Task app URL from TOML: $TASK_APP_URL"
+fi
+echo ""
+
+echo "🔍 Checking if Banking77 pipeline task app is running on ${TASK_APP_URL}..."
+if ! curl -s -f -H "X-API-Key: $ENVIRONMENT_API_KEY" "$TASK_APP_URL/health" > /dev/null 2>&1; then
+    cat <<EOF
+❌ Error: Banking77 pipeline task app is not running on ${TASK_APP_URL}
+
+Start it with:
+  uvx synth-ai deploy banking77-pipeline --runtime uvicorn --port 8112 --env-file .env --follow
+  # or deploy to Modal dev: modal deploy --env dev examples/task_apps/banking77_pipeline/deploy_wrapper.py
+EOF
+    exit 1
+fi
+echo "✅ Pipeline task app is healthy"
+echo ""
+
+echo "🔍 Checking backend connection to $BACKEND_URL..."
+if ! curl -s -f "$BACKEND_URL_NO_API/api/health" > /dev/null 2>&1; then
+    echo "⚠️  Warning: Cannot connect to backend at $BACKEND_URL"
+    read -p "Continue anyway? (y/N) " -n 1 -r
+    echo
+    if [[ ! $REPLY =~ ^[Yy]$ ]]; then
+        exit 1
+    fi
+else
+    echo "✅ Backend is healthy"
+fi
+echo ""
+
+echo "🚀 Starting MIPROv2 training with gpt-4.1-mini as policy model..."
+echo "Config: $CONFIG_PATH"
+echo ""
+echo "Multi-Step Flow:"
+echo "  1. Bootstrap: two-module pipeline on seeds [0-14]"
+echo "  2. Optimisation: 5 iterations × 2 variants (classifier + calibrator each evaluation)"
+echo "  3. Held-out evaluation on seeds [40-49]"
+echo "  Policy model: gpt-4.1-mini (OpenAI)"
+echo "  Meta-model: gpt-4o-mini (OpenAI)"
+echo ""
+
+export BACKEND_BASE_URL="$BACKEND_URL"
+export SYNTH_BASE_URL="$BACKEND_URL"
+
+uvx synth-ai train \
+    --type prompt_learning \
+    --config "$CONFIG_PATH" \
+    --backend "$BACKEND_URL" \
+    --poll
+
+echo ""
+echo "✅ MIPROv2 pipeline training complete!"
+

examples/blog_posts/mipro/verify_banking77_setup.sh

@@ -0,0 +1,117 @@
+#!/bin/bash
+# Verify Banking77 MIPROv2 setup
+
+set -e
+
+echo "🔍 Verifying Banking77 MIPROv2 Setup"
+echo "====================================="
+echo ""
+
+# Check environment variables
+echo "📋 Environment Variables:"
+MISSING_VARS=0
+
+if [ -z "$SYNTH_API_KEY" ]; then
+    echo "  ❌ SYNTH_API_KEY: not set"
+    MISSING_VARS=$((MISSING_VARS + 1))
+else
+    echo "  ✅ SYNTH_API_KEY: ${SYNTH_API_KEY:0:20}..."
+fi
+
+if [ -z "$GROQ_API_KEY" ]; then
+    echo "  ❌ GROQ_API_KEY: not set"
+    MISSING_VARS=$((MISSING_VARS + 1))
+else
+    echo "  ✅ GROQ_API_KEY: ${GROQ_API_KEY:0:20}..."
+fi
+
+if [ -z "$OPENAI_API_KEY" ]; then
+    echo "  ⚠️  OPENAI_API_KEY: not set (required for meta-model)"
+    MISSING_VARS=$((MISSING_VARS + 1))
+else
+    echo "  ✅ OPENAI_API_KEY: ${OPENAI_API_KEY:0:20}..."
+fi
+
+if [ -z "$ENVIRONMENT_API_KEY" ]; then
+    echo "  ❌ ENVIRONMENT_API_KEY: not set"
+    MISSING_VARS=$((MISSING_VARS + 1))
+else
+    echo "  ✅ ENVIRONMENT_API_KEY: ${ENVIRONMENT_API_KEY:0:20}..."
+fi
+
+echo ""
+
+# Check backend
+echo "🌐 Backend Connection:"
+BACKEND_URL="${BACKEND_BASE_URL:-http://localhost:8000}"
+if curl -s -f "$BACKEND_URL/api/health" > /dev/null 2>&1; then
+    echo "  ✅ Backend is healthy at $BACKEND_URL"
+else
+    echo "  ❌ Cannot connect to backend at $BACKEND_URL"
+    MISSING_VARS=$((MISSING_VARS + 1))
+fi
+echo ""
+
+# Check task app
+echo "📱 Task App Connection:"
+if curl -s -f -H "X-API-Key: ${ENVIRONMENT_API_KEY:-dummy}" http://127.0.0.1:8102/health > /dev/null 2>&1; then
+    echo "  ✅ Banking77 task app (single-step) is running on http://127.0.0.1:8102"
+else
+    echo "  ❌ Banking77 task app is not running on http://127.0.0.1:8102"
+    echo "     Start it with: ./examples/blog_posts/mipro/deploy_banking77_task_app.sh"
+    MISSING_VARS=$((MISSING_VARS + 1))
+fi
+
+if curl -s -f -H "X-API-Key: ${ENVIRONMENT_API_KEY:-dummy}" http://127.0.0.1:8112/health > /dev/null 2>&1; then
+    echo "  ✅ Banking77 pipeline task app is running on http://127.0.0.1:8112"
+else
+    echo "  ❌ Banking77 pipeline task app is not running on http://127.0.0.1:8112"
+    echo "     Start it with: ./examples/blog_posts/mipro/deploy_banking77_pipeline_task_app.sh"
+    MISSING_VARS=$((MISSING_VARS + 1))
+fi
+echo ""
+
+# Check config files
+echo "📄 Configuration Files:"
+CONFIG_DIR="examples/blog_posts/mipro/configs"
+if [ -f "$CONFIG_DIR/banking77_mipro_local.toml" ]; then
+    echo "  ✅ banking77_mipro_local.toml exists"
+else
+    echo "  ❌ banking77_mipro_local.toml not found"
+    MISSING_VARS=$((MISSING_VARS + 1))
+fi
+
+if [ -f "$CONFIG_DIR/banking77_mipro_test.toml" ]; then
+    echo "  ✅ banking77_mipro_test.toml exists"
+else
+    echo "  ❌ banking77_mipro_test.toml not found"
+    MISSING_VARS=$((MISSING_VARS + 1))
+fi
+
+if [ -f "$CONFIG_DIR/banking77_pipeline_mipro_local.toml" ]; then
+    echo "  ✅ banking77_pipeline_mipro_local.toml exists"
+else
+    echo "  ❌ banking77_pipeline_mipro_local.toml not found"
+    MISSING_VARS=$((MISSING_VARS + 1))
+fi
+
+if [ -f "$CONFIG_DIR/banking77_pipeline_mipro_test.toml" ]; then
+    echo "  ✅ banking77_pipeline_mipro_test.toml exists"
+else
+    echo "  ❌ banking77_pipeline_mipro_test.toml not found"
+    MISSING_VARS=$((MISSING_VARS + 1))
+fi
+echo ""
+
+# Summary
+if [ $MISSING_VARS -eq 0 ]; then
+    echo "✅ All checks passed! Ready to run MIPROv2 optimization."
+    echo ""
+    echo "Next step:"
+    echo "  ./examples/blog_posts/mipro/run_mipro_banking77.sh"
+    exit 0
+else
+    echo "❌ Setup incomplete. Please fix the issues above."
+    exit 1
+fi
+

examples/blog_posts/pokemon_vl/README.md

@@ -0,0 +1,98 @@
+# Pokémon VL: Vision-Language RL Pipeline
+
+This playbook demonstrates end-to-end vision-language reinforcement learning on Pokémon Red using Synth AI's CLI tools. We follow the eval → collect data → SFT → RL → eval pipeline, but with vision models throughout.
+
+## Overview
+
+**Model**: Qwen3-VL-4B-Instruct (4B parameter vision-language model via Synth API)
+**Environment**: Pokémon Red (Game Boy emulation with vision support)
+**Benchmark**: Pallet Town progression task (leave bedroom → get starter → win first battle)
+
+## Pipeline Steps
+
+1. **Deploy Task App** - Host the Pokémon Red environment
+2. **Collect Vision Rollouts** - Generate high-quality demonstrations using Qwen3-VL
+3. **Filter Dataset** - Extract successful trajectories for supervised fine-tuning
+4. **Fine-Tune Qwen3-4B VL** - Train vision-language model on filtered data
+5. **Vision-Language RL** - Bootstrap RL training from SFT checkpoint
+6. **Final Evaluation** - Compare SFT and RL performance
+
+## Prerequisites
+
+```bash
+# Install dependencies
+uv pip install -e .
+
+# Setup authentication
+uvx synth-ai setup
+
+# Copy environment template
+cp examples/blog_posts/pokemon_vl/.env.example .env
+```
+
+## Quick Start
+
+```bash
+# Export trace database path
+export POKEMON_VL_TRACE_DB=traces/v3/pokemon_vl_blog.db
+
+# 1. Deploy task app
+uvx synth-ai deploy pokemon_red --runtime modal --name pokemon-vl-blog --env-file .env
+
+# 2. Collect vision rollouts with Qwen3-VL
+uvx synth-ai eval pokemon_red --config examples/blog_posts/pokemon_vl/configs/eval_qwen3_vl.toml --trace-db "${POKEMON_VL_TRACE_DB}"
+
+# 3. Filter high-reward trajectories
+uvx synth-ai filter --config examples/blog_posts/pokemon_vl/configs/filter_high_reward.toml
+
+# 4. Fine-tune Qwen3-4B VL
+uvx synth-ai train --type sft --config examples/blog_posts/pokemon_vl/configs/train_sft_qwen4b_vl.toml --env-file .env --poll
+
+# 5. RL from SFT checkpoint (replace JOB_ID)
+uvx synth-ai train --type rl --config examples/blog_posts/pokemon_vl/configs/train_rl_from_sft.toml --env-file .env --poll
+
+# 6. Evaluate final RL model
+uvx synth-ai eval pokemon_red --config examples/blog_posts/pokemon_vl/configs/eval_rl_final.toml --trace-db "${POKEMON_VL_TRACE_DB}"
+```
+
+## Vision Features
+
+- **Full Game Boy Frames**: Base64-encoded PNG screenshots (160x144 resolution)
+- **Vision-Only Mode**: Pure image understanding without text state
+- **Vision + Text Mode**: Combined visual and structured state information
+- **Efficient Action Batching**: `execute_sequence` tool for 5-10 actions per inference call
+
+## Expected Results
+
+| Stage | Model | Mean Reward | Success Rate | Best Achievement |
+|-------|-------|-------------|--------------|------------------|
+| Initial | Qwen3-VL (vision) | ~150 | 60% | Win first battle |
+| SFT | Qwen3-4B VL | ~200 | 75% | Win first battle + explore |
+| RL | Qwen3-4B VL + RL | ~350 | 85% | Complete Pallet Town |
+
+## Files
+
+- `configs/` - All TOML configuration files
+- `ft_data/` - Filtered datasets for fine-tuning
+- `.env.example` - Environment variables template
+
+## Vision Model Configuration
+
+The vision models receive:
+- **Input**: Game Boy screenshot + optional structured state (position, HP, party, etc.)
+- **Output**: Sequence of button presses via `execute_sequence` tool
+- **Action Space**: UP, DOWN, LEFT, RIGHT, A, B, START, SELECT with frame counts
+
+## Reward Function
+
+Dense rewards for Pallet Town progression:
+- Leave bedroom (+20)
+- Exit house (+30)
+- Find Oak's lab (+40)
+- Talk to Oak (+50)
+- Get starter Pokémon (+100)
+- Enter battle (+75)
+- Deal damage (+50 per 10HP)
+- Win battle (+150)
+
+Total possible: ~700 points

examples/blog_posts/pokemon_vl/configs/eval_gpt5nano.toml

@@ -0,0 +1,26 @@
+[eval]
+app_id = "pokemon_red"
+task_app_url = "http://127.0.0.1:8914"
+model = "gpt-5-nano"
+seeds = [0]  # Single seed for testing
+max_turns = 10  # 10 LLM calls per episode to allow more progress
+concurrency = 1  # Run 1 rollout
+env_name = "pokemon_red"
+policy_name = "pokemon_vl_qwen3_vl"  # Reuse policy config, will override model
+trace_format = "full"
+return_trace = true
+
+[eval.policy_config]
+provider = "openai"  # Use OpenAI API for gpt-5-nano
+model = "gpt-5-nano"
+inference_url = "https://api.openai.com/v1"
+temperature = 0.7
+top_p = 0.95
+max_tokens = 512
+use_vision = true
+image_only_mode = false
+max_llm_calls = 10
+
+[eval.env_config.env_params]
+max_steps_per_episode = 100  # Allow time to achieve milestones
+

examples/blog_posts/pokemon_vl/configs/eval_qwen3_vl.toml

@@ -0,0 +1,27 @@
+[eval]
+app_id = "pokemon_red"
+task_app_url = "http://127.0.0.1:8914"
+model = "Qwen/Qwen3-VL-30B-A3B-Thinking"  # Larger thinking variant - needs more time to load
+seeds = [10, 11]  # 2 seeds for quick testing
+max_turns = 10  # 10 LLM calls per episode to allow more progress
+concurrency = 2  # Run 2 rollouts in parallel
+env_name = "pokemon_red"
+policy_name = "pokemon_vl_qwen3_vl"
+trace_format = "full"
+return_trace = true
+
+[eval.policy_config]
+provider = "synth"  # Use Synth internal API for vision models
+model = "Qwen/Qwen3-VL-30B-A3B-Thinking"  # Larger thinking variant - needs more time to load
+inference_url = "https://synth-laboratories-dev--learning-v2-service-fastapi-app.modal.run/chat/completions"
+temperature = 1.0  # Higher temperature to encourage exploration
+top_p = 0.95
+max_tokens = 2048  # Reduced to avoid token budget issues
+use_vision = true
+image_only_mode = false
+max_llm_calls = 10
+thinking_mode = "think"  # Enable thinking/reasoning mode
+thinking_budget = 3072  # Increased token budget for reasoning
+
+[eval.env_config.env_params]
+max_steps_per_episode = 100  # Increased from 3 to allow time to achieve milestones

examples/blog_posts/pokemon_vl/configs/eval_rl_final.toml

@@ -0,0 +1,24 @@
+[eval]
+app_id = "pokemon_red"
+task_app_url = "http://127.0.0.1:8914"
+model = "fft:REPLACE-WITH-RL-JOB-ID"  # Update with final RL job ID
+seeds = [100, 101, 102, 103, 104, 105, 106, 107, 108, 109]
+max_turns = 15  # Allow more steps for trained model
+concurrency = 3
+env_name = "pokemon_red"
+policy_name = "pokemon_vl_rl_final"
+trace_format = "full"
+return_trace = true
+
+[eval.policy_config]
+provider = "synth"
+model = "fft:REPLACE-WITH-RL-JOB-ID"  # Update with final RL job ID
+temperature = 0.1  # Lower temperature for evaluation
+top_p = 0.9
+max_tokens = 4096
+use_vision = true
+image_only_mode = false
+max_llm_calls = 15
+
+[eval.env_config.env_params]
+max_steps_per_episode = 15

examples/blog_posts/pokemon_vl/configs/filter_high_reward.toml

@@ -0,0 +1,10 @@
+# Filter high-quality vision-language rollouts for SFT training
+# Assumes traces stored in pokemon_vl_blog.db via eval commands
+
+[filter]
+db = "traces/v3/pokemon_vl_blog.db"
+output = "examples/blog_posts/pokemon_vl/ft_data/pokemon_vl_high_reward.jsonl"
+min_official_score = 0.3  # Require at least 30% completion (Pallet Town progression)
+models = ["Qwen/Qwen3-VL-4B-Instruct"]  # Vision models used for rollouts
+shuffle = true
+shuffle_seed = 42

examples/blog_posts/pokemon_vl/configs/train_rl_from_sft.toml

@@ -0,0 +1,43 @@
+# Vision-Language RL: Continue training Qwen3-4B VL from SFT checkpoint
+# Update task_url with deployed Modal task app URL
+# Set model.source to the SFT job id from `uvx synth-ai train --type sft`
+
+type = "rl"
+
+[services]
+task_url = "http://127.0.0.1:8914"
+
+[compute]
+gpu_type = "H100"
+gpu_count = 8
+
+[topology]
+gpus_for_vllm = 4
+gpus_for_training = 3
+gpus_for_ref = 1
+
+[vllm]
+tensor_parallel_size = 4
+
+[model]
+source = "fft:REPLACE-WITH-SFT-JOB-ID"  # Update with actual SFT job ID
+label = "pokemon_vl_rl_blog"
+supports_vision = true
+
+[rollout]
+max_turns = 10
+episodes_per_batch = 64
+task_app_origin_rewards_only = true
+
+[evaluation]
+instances = 100
+every_n_iters = 20
+seeds = [0, 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]
+
+[training]
+log_interval = 1
+
+[training.weight_sync]
+enable = true
+targets = ["policy"]
+weight_sync_interval = 1

examples/blog_posts/pokemon_vl/configs/train_sft_qwen4b_vl.toml

@@ -0,0 +1,40 @@
+# Vision-Language Supervised Fine-Tuning: Qwen3-4B VL on filtered Pokémon rollouts
+# Update the `data` path once `uvx synth-ai filter` produces your JSONL
+
+[algorithm]
+type = "offline"
+method = "sft"
+variety = "fft"
+
+[job]
+model = "Qwen/Qwen3-VL-4B-Instruct"  # Vision-enabled Qwen3-VL model
+data = "../ft_data/pokemon_vl_high_reward.jsonl"
+poll_seconds = 1800
+
+[compute]
+gpu_type = "H100"
+gpu_count = 4
+nodes = 1
+
+[data.topology]
+container_count = 4
+
+[training]
+mode = "full_finetune"
+use_qlora = false
+
+[hyperparameters]
+n_epochs = 2
+world_size = 4
+sequence_length = 4096  # Longer for vision tokens + text
+per_device_batch = 2
+gradient_accumulation_steps = 64
+learning_rate = 8e-6
+warmup_ratio = 0.03
+
+[hyperparameters.parallelism]
+use_deepspeed = true
+deepspeed_stage = 3
+fsdp = false
+bf16 = true
+fp16 = false