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

examples/qwen_vl/NEXT_STEPS_2B.md

@@ -0,0 +1,325 @@
+# Next Steps: Qwen3-VL-2B SFT & RL Training
+
+**Status:** Data collection complete ✅ | Ready for SFT training 🚀
+
+---
+
+## 📋 Current Status
+
+### ✅ Completed
+1. **VLM Data Collection Pipeline** - WORKING END-TO-END
+   - Fixed task app tracing to return full session traces
+   - Fixed CLI to handle multimodal content preservation
+   - Successfully collected traces with base64 PNG images
+   - Database: `traces/gpt4o_vision_test/rollouts.db`
+   - Exported: `traces/gpt4o_vision_test/sft/train.jsonl` (50 examples validated)
+
+2. **Infrastructure Validated**
+   - ✅ `synth-ai eval` stores traces with images
+   - ✅ `synth-ai filter` exports SFT JSONL with preserved images
+   - ✅ Multimodal messages follow OpenAI format
+   - ✅ Images embedded as base64 PNG (~1306 chars per 64x64 image)
+
+3. **Documentation**
+   - `VLM_PIPELINE_COMPLETE.md` - Full pipeline guide
+   - `PIPELINE_RUN_LOG.txt` - Execution log with all fixes
+   - `BUGS_AND_FIXES.md` - Detailed bug reports
+   - `SETUP_COMPLETE.md` - Summary of setup
+
+---
+
+## 🎯 Next Steps: Train Qwen3-VL-2B
+
+### Step 1: Scale Up Data Collection (Optional)
+
+We have 50 working examples. For production training, collect more:
+
+```bash
+cd /Users/joshpurtell/Documents/GitHub/synth-ai
+
+# Collect 100 episodes (will create ~5000 samples)
+export TASKAPP_TRACING_ENABLED=1
+uvx synth-ai eval \
+  --config examples/qwen_vl/configs/eval_gpt4o_vision_proper.toml \
+  --seeds 0-99 \
+  --trace-db traces/gpt4o_vision_100/rollouts.db \
+  --env-file /path/to/.env
+
+# Filter and export
+uvx synth-ai filter \
+  --config examples/qwen_vl/configs/filter_vision_test.toml
+```
+
+**Output:** ~4500 SFT examples with images
+
+---
+
+### Step 2: Create SFT Config for Qwen3-VL-2B
+
+File: `/Users/joshpurtell/Documents/GitHub/monorepo/configs/vision_sft/crafter_qwen3vl_2b_gpt4o.toml`
+
+```toml
+# Crafter Vision SFT: Qwen3-VL-2B trained on gpt-4o-mini traces
+# Using 2B model for faster iteration and lower GPU requirements
+
+[algorithm]
+type = "offline"
+method = "sft"
+variety = "lora"
+
+[job]
+model = "Qwen/Qwen3-VL-2B-Instruct"
+data = "traces/gpt4o_vision_100/sft/train.jsonl"
+
+[compute]
+gpu_type = "H200"
+gpu_count = 2
+nodes = 1
+
+[training]
+mode = "lora"
+use_qlora = true
+
+[hyperparameters]
+n_epochs = 3
+per_device_batch = 1
+gradient_accumulation_steps = 16
+sequence_length = 2048
+learning_rate = 5e-05
+warmup_ratio = 0.03
+train_kind = "peft"
+
+# LoRA config
+lora_rank = 16
+lora_alpha = 32
+lora_dropout = 0.05
+lora_target_modules = ["all-linear"]
+
+# Training optimizations
+[hyperparameters.parallelism]
+use_deepspeed = true
+deepspeed_stage = 2
+bf16 = true
+activation_checkpointing = true
+
+# Evaluation
+evaluation_strategy = "steps"
+eval_steps = 50
+save_best_model_at_end = true
+metric_for_best_model = "val.loss"
+
+[tags]
+task = "crafter"
+modality = "vision"
+data_source = "openai_gpt4o_mini"
+model_family = "qwen3_vl"
+model_size = "2b"
+```
+
+---
+
+### Step 3: Run SFT Training
+
+```bash
+cd /Users/joshpurtell/Documents/GitHub/monorepo
+
+# Copy data to monorepo (if not already there)
+cp -r /Users/joshpurtell/Documents/GitHub/synth-ai/traces/gpt4o_vision_100/sft/ \
+   backend/data/vision_sft/
+
+# Submit SFT job
+export BACKEND_BASE_URL="https://synth-backend-dev-docker.onrender.com/api"
+uvx synth-ai train \
+  --type sft \
+  --config configs/vision_sft/crafter_qwen3vl_2b_gpt4o.toml \
+  --env-file backend/.env.dev
+```
+
+**Expected:**
+- Training time: 1-2 hours
+- Cost: ~$10.50 (2x H200)
+- Output: LoRA adapter at `lora_adapters/qwen3vl_2b_crafter_gpt4o/`
+
+---
+
+### Step 4: Create RL Config for Qwen3-VL-2B
+
+File: `/Users/joshpurtell/Documents/GitHub/synth-ai/examples/qwen_vl/configs/crafter_rl_qwen3vl_2b.toml`
+
+```toml
+# Crafter Vision RL: Qwen3-VL-2B with GRPO
+# Uses SFT-initialized model for RL fine-tuning
+
+[algorithm]
+type = "online"
+method = "grpo"
+variety = "default"
+
+[model]
+base = "Qwen/Qwen3-VL-2B-Instruct"
+adapter = "lora_adapters/qwen3vl_2b_crafter_gpt4o"  # From SFT step
+
+[job]
+rollout_count = 50
+n_iterations = 20
+max_steps_per_rollout = 50
+
+[compute]
+gpu_type = "H200"
+gpu_count = 4
+nodes = 1
+
+[topology]
+type = "single_node_split"
+gpus_for_vllm = 2
+gpus_for_training = 2
+gpus_for_ref = 0
+
+[vllm]
+tensor_parallel_size = 1  # 2B fits on 1 GPU
+enable_prefix_caching = false
+use_cudagraph = true
+gpu_memory_utilization = 0.85
+max_model_len = 2048
+
+[training]
+mode = "lora"
+use_qlora = true
+
+[hyperparameters]
+per_device_batch = 2
+gradient_accumulation_steps = 8
+sequence_length = 2048
+learning_rate = 2e-06
+warmup_ratio = 0.1
+train_kind = "peft"
+
+lora_rank = 16
+lora_alpha = 32
+lora_dropout = 0.05
+lora_target_modules = ["all-linear"]
+
+[grpo]
+kl_coeff = 0.1
+clip_range = 0.2
+value_clip_range = 0.2
+normalize_rewards = true
+
+[judge]
+type = "remote"
+provider = "openai"
+model = "gpt-4o-mini"
+
+[tags]
+task = "crafter"
+modality = "vision"
+algorithm = "grpo"
+model_family = "qwen3_vl"
+model_size = "2b"
+```
+
+---
+
+### Step 5: Run RL Training
+
+```bash
+cd /Users/joshpurtell/Documents/GitHub/synth-ai
+
+# Submit RL job
+uvx synth-ai train \
+  --type rl \
+  --config examples/qwen_vl/configs/crafter_rl_qwen3vl_2b.toml \
+  --env-file /path/to/.env
+```
+
+**Expected:**
+- Training time: 4-6 hours
+- Cost: ~$70 (4x H200)
+- Output: RL-tuned adapter at `lora_adapters/qwen3vl_2b_crafter_rl_iter20/`
+
+---
+
+### Step 6: Evaluate Results
+
+```bash
+# Run benchmark
+python examples/qwen_vl/benchmark_vision_agents.py
+```
+
+**Expected Performance:**
+- Base Qwen3-VL-2B: ~6.5% achievement rate
+- After SFT: ~20% achievement rate (+13.5%)
+- After RL: ~38% achievement rate (+18% more)
+- Teacher (gpt-4o-mini): ~45% achievement rate
+
+---
+
+## 💰 Cost & Timeline Summary
+
+### Qwen3-VL-2B Pipeline
+
+| Step | Description | Cost | Time |
+|------|-------------|------|------|
+| 1 | Data collection (100 episodes) | ~$1-2 | 30-60 min |
+| 2 | Dataset assembly | $0 | < 5 min |
+| 3 | Vision SFT (3 epochs) | ~$10.50 | 1-2 hrs |
+| 4 | Vision RL (20 iterations) | ~$70 | 4-6 hrs |
+| 5 | Evaluation | ~$5 | 2-3 hrs |
+
+**Total:** ~$87, 8-12 hours
+
+### Cost Comparison: 2B vs 8B
+
+| Model | SFT Cost | RL Cost | Total | Training Time |
+|-------|----------|---------|-------|---------------|
+| 2B    | $10.50   | $70     | $87   | 8-12 hrs      |
+| 8B    | $21      | $112    | $140  | 12-18 hrs     |
+
+**Savings with 2B:** 40% cost reduction, 30% faster
+
+---
+
+## 🎯 Key Advantages of 2B Model
+
+1. **Faster Iteration**
+   - SFT: 1-2 hours vs 2-4 hours for 8B
+   - RL: 4-6 hours vs 6-10 hours for 8B
+   - Enables rapid experimentation
+
+2. **Lower GPU Requirements**
+   - Fits on 1 GPU for inference (use 2 for safety)
+   - Can use batch_size=2 vs 1 for 8B
+   - More efficient GPU utilization
+
+3. **Cost Effective**
+   - ~$87 total vs $140 for 8B
+   - Better for initial prototyping
+   - Scale to 8B later if needed
+
+4. **Competitive Performance**
+   - ~38% achievement rate after RL
+   - vs ~42% for 8B (only 4% difference)
+   - Good enough for validation and testing
+
+---
+
+## 📝 Notes
+
+- All configs use LoRA for memory efficiency
+- Vision models require batch_size=1-2 (images are memory-intensive)
+- Use DeepSpeed Stage 2 for training optimization
+- Disable prefix caching (unstable with LoRA + vision)
+- 2B model is perfect for initial testing and rapid iteration
+
+---
+
+## 🚀 Ready to Start!
+
+The infrastructure is ready. Just need to:
+1. Create the two TOML configs above
+2. Run SFT training
+3. Run RL training
+4. Evaluate and compare
+
+All the hard work (data collection, tracing fixes, filtering) is done! 🎉
+

examples/qwen_vl/QUICKSTART.md

@@ -0,0 +1,327 @@
+# Qwen VL Quickstart Guide
+
+Complete guide to running vision-language models on Crafter with image observations.
+
+## 🚀 Quick Demo
+
+### Option 1: Run gpt-5-nano (OpenAI)
+```bash
+export OPENAI_API_KEY="sk-..."
+uv run python examples/qwen_vl/crafter_gpt5nano_agent.py --seeds 5 --steps 10
+```
+
+### Option 2: Run Qwen-VL (synth-ai)
+```bash
+export SYNTH_API_KEY="sk_live_..."
+uv run python examples/qwen_vl/crafter_qwen_vl_agent.py \
+  --model Qwen/Qwen2-VL-7B-Instruct --seeds 5 --steps 10
+```
+
+### Option 3: Compare Both
+```bash
+export OPENAI_API_KEY="sk-..."
+export SYNTH_API_KEY="sk_live_..."
+bash examples/qwen_vl/run_vision_comparison.sh
+```
+
+---
+
+## 📊 Expected Output
+
+```
+Running 10 Crafter episodes with model=gpt-5-nano
+Using OpenAI API
+
+Seed 00: steps=10, achievements=2, tool_calls=10, reward≈1.250
+Seed 01: steps=10, achievements=1, tool_calls=10, reward≈0.750
+Seed 02: steps=10, achievements=3, tool_calls=10, reward≈1.500
+...
+
+Summary
+-------
+{
+  "model": "gpt-5-nano",
+  "provider": "openai",
+  "episodes": 10,
+  "mean_steps": 9.8,
+  "mean_achievements": 2.1,
+  "total_tool_calls": 98,
+  "output_dir": "examples/qwen_vl/temp/gpt5nano_frames"
+}
+
+Frames saved in: examples/qwen_vl/temp/gpt5nano_frames/
+```
+
+Each episode saves PNG frames (64x64) showing what the VLM saw:
+```
+examples/qwen_vl/temp/gpt5nano_frames/
+  seed_0000/
+    step_000.png
+    step_001.png
+    step_002.png
+    ...
+  seed_0001/
+    ...
+```
+
+---
+
+## 🎯 Full Pipeline: Data Collection → SFT → RL
+
+### Step 1: Collect Vision Traces
+
+Collect 100 episodes with gpt-5-nano (for teacher distillation):
+
+```bash
+export OPENAI_API_KEY="sk-..."
+
+uv run python examples/qwen_vl/collect_vision_traces.py \
+  --model gpt-5-nano \
+  --provider openai \
+  --episodes 100 \
+  --max-steps 50 \
+  --output-dir traces/gpt5nano_vision
+```
+
+**Output:**
+- SQLite DB: `traces/gpt5nano_vision/rollouts.db`
+- Contains multimodal traces with images
+- ~5000 samples (100 episodes × ~50 steps)
+
+**Timeline:** 30-60 minutes  
+**Cost:** ~$1-2 (OpenAI gpt-5-nano)
+
+---
+
+### Step 2: Export to SFT JSONL
+
+Convert SQLite traces to SFT training format:
+
+```bash
+uv run python examples/qwen_vl/export_traces_to_sft.py \
+  --db-path traces/gpt5nano_vision/rollouts.db \
+  --output traces/gpt5nano_vision/sft_dataset.jsonl \
+  --min-steps 5
+```
+
+**Output:**
+- JSONL file with OpenAI-format messages
+- Each line: `{"messages": [...], "metadata": {...}}`
+- Messages include base64-encoded images
+
+---
+
+### Step 3: Split Train/Val
+
+```bash
+uv run python examples/qwen_vl/split_sft_data.py \
+  --input traces/gpt5nano_vision/sft_dataset.jsonl \
+  --train-output traces/gpt5nano_vision/train.jsonl \
+  --val-output traces/gpt5nano_vision/val.jsonl \
+  --val-fraction 0.1
+```
+
+**Output:**
+- `train.jsonl`: ~4400 samples
+- `val.jsonl`: ~500 samples
+
+---
+
+### Step 4: Train Vision SFT
+
+Use the example config or create your own:
+
+```bash
+cd /path/to/monorepo
+
+export BACKEND_BASE_URL="https://synth-backend-dev-docker.onrender.com/api"
+
+uvx synth-ai train \
+  --type sft \
+  --config examples/qwen_vl/configs/crafter_vlm_sft_example.toml \
+  --env-file backend/.env.dev
+```
+
+**Hardware:** 2x H200 (or 4x H100)  
+**Time:** 2-4 hours (2 epochs)  
+**Cost:** ~$21 (Modal GPU pricing)
+
+**Output:**
+- LoRA adapter saved to HF Hub or S3
+- Wandb logs with training curves
+
+---
+
+### Step 5: Run Vision RL (Optional)
+
+After SFT, fine-tune with GRPO for better performance:
+
+```toml
+# example RL config
+[algorithm]
+type = "online"
+method = "grpo"
+
+[model]
+base = "Qwen/Qwen2-VL-7B-Instruct"
+adapter = "s3://my-bucket/qwen2vl_crafter_sft"  # From SFT
+
+[compute]
+gpu_count = 4  # 2 inference + 2 training
+```
+
+**Time:** 6-10 hours (20 iterations)  
+**Cost:** ~$112
+
+---
+
+## 📁 File Structure
+
+```
+synth-ai/examples/qwen_vl/
+├── README.md                       # Overview
+├── QUICKSTART.md                   # This file
+├── __init__.py
+│
+├── crafter_gpt5nano_agent.py       # OpenAI gpt-5-nano demo
+├── crafter_qwen_vl_agent.py        # Qwen-VL (synth-ai) demo
+├── collect_vision_traces.py        # Trace collection for SFT
+├── run_vision_comparison.sh        # Compare both models
+│
+├── configs/
+│   └── crafter_vlm_sft_example.toml  # Example SFT config
+│
+└── temp/                           # Output frames and summaries
+    ├── gpt5nano_frames/
+    ├── qwen_vl_frames/
+    └── comparison/
+```
+
+---
+
+## 🔍 How Vision Detection Works
+
+CrafterPolicy automatically detects vision capability:
+
+```python
+# From examples/task_apps/crafter/.../policy.py
+@staticmethod
+def _is_vision_model(model_name: str) -> bool:
+    """Check if model supports vision from its name."""
+    model_lower = model_name.lower()
+    
+    vision_patterns = [
+        "gpt-5",           # ✅ gpt-5-nano, gpt-5-turbo, etc.
+        "gpt-4o",          # ✅ gpt-4o-mini, gpt-4o
+        "qwen-vl",         # ✅ Qwen-VL-Chat
+        "qwen2-vl",        # ✅ Qwen2-VL-7B-Instruct
+        "qwen3-vl",        # ✅ Qwen3-VL-8B
+        # ... more patterns
+    ]
+    
+    return any(pattern in model_lower for pattern in vision_patterns)
+```
+
+If detected:
+- Policy includes base64 image in user message
+- Images are 64x64 PNG frames from Crafter
+- Format: `{"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}}`
+
+---
+
+## 🎛️ Advanced Configuration
+
+### Custom Image Resolution
+
+Edit Crafter task instance config:
+
+```python
+instance.config = {
+    "seed": seed,
+    "length": 256,
+    "area": [128, 128],  # Higher resolution (default: 64x64)
+}
+```
+
+**Note:** Higher resolution = more tokens = higher cost
+
+### Image-Only Mode
+
+Disable text observations, use only images:
+
+```python
+await policy.initialize({
+    "use_tools": True,
+    "model": model,
+    "image_only_mode": True,  # No text, only images
+})
+```
+
+### Multiple Images per Step
+
+For temporal context (not yet implemented):
+
+```python
+# Future: Include last N frames
+image_parts = [
+    {"type": "image_url", "image_url": {"url": frame_t}},
+    {"type": "image_url", "image_url": {"url": frame_t_minus_1}},
+    {"type": "image_url", "image_url": {"url": frame_t_minus_2}},
+]
+```
+
+---
+
+## 🐛 Troubleshooting
+
+### Error: `OPENAI_API_KEY not set`
+```bash
+export OPENAI_API_KEY="sk-..."
+```
+
+### Error: `SYNTH_API_KEY not set`
+```bash
+export SYNTH_API_KEY="sk_live_..."
+```
+
+### Error: `TracingStore not available`
+Traces require synth-ai tracing module:
+```bash
+uv sync  # Ensure all dependencies are installed
+```
+
+### Vision not detected
+Manually enable:
+```python
+await policy.initialize({"use_vision": True})
+```
+
+---
+
+## 📚 Related Documentation
+
+- **SFT Pipeline:** See `/Users/joshpurtell/Documents/GitHub/monorepo/vision_sft_rl.txt` (Phase 9)
+- **Crafter Environment:** `examples/task_apps/crafter/README.md`
+- **OpenAI VLM Examples:** `examples/vlm/crafter_openai_vlm_agent.py`
+- **Image-Only Eval:** `examples/task_apps/IMAGE_ONLY_EVAL_QUICKSTART.md`
+
+---
+
+## 🎉 Next Steps
+
+1. ✅ Run demos to verify vision inference works
+2. 🎯 Collect training traces (100-1000 episodes)
+3. 📦 Export and split into train/val
+4. 🚀 Train VLM with LoRA (see `crafter_vlm_sft_example.toml`)
+5. 🏆 Fine-tune with RL/GRPO for better achievement rates
+6. 📊 Benchmark: base model vs SFT vs SFT+RL
+
+**Expected Performance:**
+- Base Qwen-VL: ~5-10% achievement rate
+- After SFT (gpt-5-nano distillation): ~20-30%
+- After RL (20 iterations): ~40-50%
+
+---
+
+Happy vision-language model training! 🚀✨
+