synth-ai 0.2.16__py3-none-any.whl → 0.2.17__py3-none-any.whl

examples/qwen_vl/README.md

@@ -55,21 +55,21 @@ uvx synth-ai train --type sft --config configs/vision_sft/crafter_qwen3vl_8b_gpt
 Run Crafter agent using Qwen-VL models via synth-ai's hosted inference.
 
 **Models supported:**
-- `Qwen/Qwen2-VL-7B-Instruct`
-- `Qwen/Qwen2-VL-2B-Instruct`
-- `Qwen/Qwen3-VL-8B` (or any Qwen VL variant)
+- `Qwen/Qwen3-VL-2B-Instruct`
+- `Qwen/Qwen3-VL-4B-Instruct`
+- `Qwen/Qwen3-VL-8B-Instruct` (or any Qwen3 VL variant)
 
 **Usage:**
 ```bash
-# Run with Qwen2-VL-7B
+# Run with Qwen3-VL-4B
 uv run python examples/qwen_vl/crafter_qwen_vl_agent.py \
-  --model Qwen/Qwen2-VL-7B-Instruct \
+  --model Qwen/Qwen3-VL-4B-Instruct \
   --seeds 10 \
   --steps 20
 
 # Run with Qwen3-VL-8B  
 uv run python examples/qwen_vl/crafter_qwen_vl_agent.py \
-  --model Qwen/Qwen3-VL-8B \
+  --model Qwen/Qwen3-VL-8B-Instruct \
   --seeds 10 \
   --steps 20
 ```
@@ -113,13 +113,13 @@ uv run python examples/qwen_vl/collect_vision_traces.py \
   --max-steps 50 \
   --output-dir traces/gpt5nano_vision
 
-# Collect traces with Qwen2-VL via synth
+# Collect traces with Qwen3-VL via synth
 uv run python examples/qwen_vl/collect_vision_traces.py \
-  --model Qwen/Qwen2-VL-7B-Instruct \
+  --model Qwen/Qwen3-VL-8B-Instruct \
   --provider synth \
   --episodes 100 \
   --max-steps 50 \
-  --output-dir traces/qwen2vl_vision
+  --output-dir traces/qwen3vl_vision
 ```
 
 **Output:** SQLite database with multimodal traces ready for SFT export.
@@ -132,8 +132,7 @@ CrafterPolicy automatically detects vision capability from model names:
 - ✅ `gpt-5*` → Vision enabled
 - ✅ `gpt-4o*` → Vision enabled  
 - ✅ `*qwen-vl*` → Vision enabled
-- ✅ `*qwen2-vl*` → Vision enabled
-- ✅ `qwen3-vl*` → Vision enabled
+- ✅ `*qwen3-vl*` → Vision enabled
 
 Or set explicitly: `policy.use_vision = True`
 
@@ -151,4 +150,3 @@ Crafter environment provides observations as:
 3. Export to SFT JSONL format (see `vision_sft_rl.txt`)
 4. Train VLM with LoRA (see monorepo SFT configs)
 5. Fine-tune with RL/GRPO
-

examples/qwen_vl/SETUP_COMPLETE.md

@@ -15,10 +15,10 @@ Complete vision-language model (VLM) infrastructure for Crafter with image obser
 
 ### **Configuration Files**
 6. **`configs/eval_gpt5nano_vision.toml`** - Eval config for gpt-5-nano
-7. **`configs/eval_qwen2vl_vision.toml`** - Eval config for Qwen2-VL
+7. **`configs/eval_qwen3vl_vision.toml`** - Eval config for Qwen3-VL
 8. **`configs/eval_gpt4o_mini_vision.toml`** - Eval config for gpt-4o-mini (stronger teacher)
 9. **`configs/filter_vision_sft.toml`** - Filter config for gpt-5-nano traces
-10. **`configs/filter_qwen2vl_sft.toml`** - Filter config for Qwen2-VL traces
+10. **`configs/filter_qwen3vl_sft.toml`** - Filter config for Qwen3-VL traces
 11. **`configs/crafter_vlm_sft_example.toml`** - Example SFT training config
 
 ### **Documentation**
@@ -81,7 +81,7 @@ uv run python examples/qwen_vl/crafter_gpt5nano_agent.py --seeds 5 --steps 10
 - Stores traces to SQLite with base64-encoded images
 - Supports parallel episodes for faster collection
 
-**Config:** `eval_gpt5nano_vision.toml`, `eval_qwen2vl_vision.toml`, etc.
+**Config:** `eval_gpt5nano_vision.toml`, `eval_qwen3vl_vision.toml`, etc.
 
 ### **synth-ai filter** (Quality Filtering)
 - Removes low-quality episodes (too short, errors, loops)
@@ -89,7 +89,7 @@ uv run python examples/qwen_vl/crafter_gpt5nano_agent.py --seeds 5 --steps 10
 - Exports to SFT JSONL format (OpenAI-style messages)
 - Splits into train/val sets
 
-**Config:** `filter_vision_sft.toml`, `filter_qwen2vl_sft.toml`
+**Config:** `filter_vision_sft.toml`, `filter_qwen3vl_sft.toml`
 
 ### **synth-ai train** (Model Training)
 - Trains VLM with LoRA on collected traces
@@ -194,13 +194,13 @@ model = "gpt-4o-mini-2024-07-18"  # Stronger teacher
 ### Collect More Episodes
 ```toml
 [eval]
-num_episodes = 500  # Default: 100
+seeds = "0-499"  # Default: "0-99"
 ```
 
 ### Change Image Resolution
 ```toml
-[task.config]
-render_size = [128, 128]  # Default: [64, 64]
+[eval.env_config]
+env_params = {render_size = [128, 128]}  # Default: [64, 64]
 ```
 
 ### Adjust Quality Filters
@@ -272,4 +272,3 @@ min_achievements_per_episode = 0
 ---
 
 **Infrastructure ready!** 🎉 Start collecting vision traces and training your VLM! 🚀
-

examples/qwen_vl/VISION_TESTS_COMPLETE.md

@@ -33,7 +33,7 @@ test_vision_inference_multiple_images()         # Multiple images per message
 **File:** `tests/integration/cli/test_cli_train_sft_vision.py`
 
 ```python
-test_cli_train_sft_vision_qwen2vl()            # Full SFT job submission
+test_cli_train_sft_vision_qwen3vl()            # Full SFT job submission
 test_vision_sft_dataset_validation()           # Dataset quality checks
 test_cli_train_sft_vision_small_config()       # Fast CI test
 ```
@@ -478,7 +478,7 @@ tests/integration/cli/test_cli_inference_vision.py::test_vision_inference_valida
 tests/integration/cli/test_cli_inference_vision.py::test_vision_inference_multiple_images PASSED
 tests/integration/cli/test_cli_train_sft_vision.py::test_vision_sft_dataset_validation PASSED
 tests/integration/cli/test_cli_train_sft_vision.py::test_cli_train_sft_vision_small_config PASSED
-tests/integration/cli/test_cli_train_sft_vision.py::test_cli_train_sft_vision_qwen2vl PASSED
+tests/integration/cli/test_cli_train_sft_vision.py::test_cli_train_sft_vision_qwen3vl PASSED
 tests/integration/cli/test_cli_train_rl_vision.py::test_task_app_vision_support PASSED
 tests/integration/cli/test_cli_train_rl_vision.py::test_cli_train_rl_vision_small_config PASSED
 tests/integration/cli/test_cli_train_rl_vision.py::test_cli_train_rl_vision_qwen3vl4b PASSED
@@ -487,4 +487,3 @@ tests/integration/cli/test_cli_train_rl_vision.py::test_cli_train_rl_vision_qwen
 ```
 
 **Status:** 🎯 Production-ready! Complete vision ML pipeline tested from inference through RL training! 🎉
-

examples/qwen_vl/collect_data_via_cli.md

@@ -5,8 +5,8 @@ Use synth-ai's built-in CLI tools to collect vision traces for SFT training.
 ## 📋 Overview
 
 **Pipeline:**
-1. `synth-ai serve` → Start Crafter task app with vision support
-2. `synth-ai eval` → Run rollouts with gpt-5-nano or Qwen-VL, collect traces
+1. `synth-ai deploy --runtime=uvicorn` → Start the Crafter task app locally
+2. `synth-ai eval` → Run rollouts with GPT-4o Mini or Qwen3-VL and collect traces
 3. `synth-ai filter` → Filter traces by quality, convert to SFT format
 
 ---
@@ -19,9 +19,10 @@ Use synth-ai's built-in CLI tools to collect vision traces for SFT training.
 cd /Users/joshpurtell/Documents/GitHub/synth-ai
 
 # Serve Crafter task app on localhost:8000
-uvx synth-ai serve \
-  --task-app examples/task_apps/crafter/task_app/synth_envs_hosted/main.py \
-  --port 8000
+uvx synth-ai deploy grpo-crafter-task-app \
+  --runtime uvicorn \
+  --port 8000 \
+  --trace traces/v3
 ```
 
 **Output:**
@@ -32,7 +33,7 @@ uvx synth-ai serve \
 
 ### Option B: Use Hosted Task App (Modal)
 
-If you have a deployed Crafter task app on Modal:
+If you already have a deployed Crafter task app on Modal:
 ```bash
 export TASK_APP_URL="https://synth-laboratories--grpo-crafter-task-app.modal.run"
 ```
@@ -41,40 +42,36 @@ export TASK_APP_URL="https://synth-laboratories--grpo-crafter-task-app.modal.run
 
 ## 🎯 Step 2: Run Eval with Vision Models
 
-### Collect gpt-5-nano Traces (OpenAI)
+### Collect GPT-4o-mini Vision Traces (OpenAI)
 
 Create eval config: `examples/qwen_vl/configs/eval_gpt5nano_vision.toml`
 
 ```toml
-# Evaluation config for gpt-5-nano with vision
+# Evaluation config for gpt-4o-mini (vision)
+# Legacy filename kept for convenience
 [eval]
-model = "gpt-5-nano"
-provider = "openai"  # Use OpenAI API
+app_id = "grpo-crafter-task-app"
 task_app_url = "http://localhost:8000"  # or your hosted URL
-
-# Vision settings
-use_vision = true
-image_only_mode = false  # Include both text + images
-
-# Rollout settings
-num_episodes = 100
-max_steps_per_episode = 50
-seeds = "0-99"  # Seeds 0 through 99
-
-# Sampling
+model = "gpt-4o-mini-2024-07-18"
+seeds = "0-99"
+max_turns = 50
+concurrency = 5
+env_name = "crafter"
+policy_name = "crafter-react"
+trace_format = "structured"
+return_trace = true
+
+[eval.env_config]
+env_params = {max_steps_per_episode = 50}
+
+[eval.policy_config]
+provider = "openai"
+model = "gpt-4o-mini-2024-07-18"
 temperature = 0.7
 max_tokens = 512
-
-# Trace collection
-collect_traces = true
-trace_db = "traces/gpt5nano_vision/rollouts.db"
-
-# Tools
+use_vision = true
+image_only_mode = false
 use_tools = true
-
-[task]
-name = "crafter"
-environment = "crafter-classic"
 ```
 
 **Run evaluation:**
@@ -83,15 +80,15 @@ export OPENAI_API_KEY="sk-..."
 
 uvx synth-ai eval \
   --config examples/qwen_vl/configs/eval_gpt5nano_vision.toml \
-  --output-dir traces/gpt5nano_vision
+  --trace-db traces/gpt4omini_vision/rollouts.db
 ```
 
 **Expected output:**
 ```
-🎮 Running evaluation: gpt-5-nano on crafter
+🎮 Running evaluation: gpt-4o-mini on crafter
 📊 Episodes: 100, Max steps: 50
 🔍 Vision: enabled (auto-detected from model name)
-📦 Collecting traces to: traces/gpt5nano_vision/rollouts.db
+📦 Collecting traces to: traces/gpt4omini_vision/rollouts.db
 
 Episode 0/100 (seed=0): 50 steps, 3 achievements ✓
 Episode 1/100 (seed=1): 48 steps, 2 achievements ✓
@@ -103,45 +100,40 @@ Episode 99/100 (seed=99): 50 steps, 3 achievements ✓
    Total episodes: 100
    Total steps: 4,923
    Avg achievements: 2.8
-   Traces saved to: traces/gpt5nano_vision/rollouts.db
+   Traces saved to: traces/gpt4omini_vision/rollouts.db
 ```
 
 ---
 
-### Collect Qwen-VL Traces (synth-ai hosted)
+### Collect Qwen3-VL Traces (Synth hosted inference)
 
-Create eval config: `examples/qwen_vl/configs/eval_qwen2vl_vision.toml`
+Create eval config: `examples/qwen_vl/configs/eval_qwen3vl_vision.toml`
 
 ```toml
-# Evaluation config for Qwen2-VL via synth-ai
+# Evaluation config for Qwen3-VL vision rollouts
 [eval]
-model = "Qwen/Qwen2-VL-7B-Instruct"
-provider = "synth"  # Use synth-ai hosted inference
+app_id = "grpo-crafter-task-app"
 task_app_url = "http://localhost:8000"
-
-# Vision settings (auto-detected from model name)
-use_vision = true
-image_only_mode = false
-
-# Rollout settings
-num_episodes = 100
-max_steps_per_episode = 50
-seeds = "0-99"
-
-# Sampling
+model = "Qwen/Qwen3-VL-8B-Instruct"
+seeds = "100-199"
+max_turns = 50
+concurrency = 5
+env_name = "crafter"
+policy_name = "crafter-react"
+trace_format = "structured"
+return_trace = true
+
+[eval.env_config]
+env_params = {max_steps_per_episode = 50}
+
+[eval.policy_config]
+provider = "synth"
+model = "Qwen/Qwen3-VL-8B-Instruct"
 temperature = 0.7
 max_tokens = 512
-
-# Trace collection
-collect_traces = true
-trace_db = "traces/qwen2vl_vision/rollouts.db"
-
-# Tools
+use_vision = true
+image_only_mode = false
 use_tools = true
-
-[task]
-name = "crafter"
-environment = "crafter-classic"
 ```
 
 **Run evaluation:**
@@ -149,8 +141,8 @@ environment = "crafter-classic"
 export SYNTH_API_KEY="sk_live_..."
 
 uvx synth-ai eval \
-  --config examples/qwen_vl/configs/eval_qwen2vl_vision.toml \
-  --output-dir traces/qwen2vl_vision
+  --config examples/qwen_vl/configs/eval_qwen3vl_vision.toml \
+  --trace-db traces/qwen3vl_vision/rollouts.db
 ```
 
 ---
@@ -169,8 +161,8 @@ Create `examples/qwen_vl/configs/filter_vision_sft.toml`:
 ```toml
 # Filter vision traces for SFT training
 [filter]
-input_db = "traces/gpt5nano_vision/rollouts.db"
-output_dir = "traces/gpt5nano_vision/sft"
+input_db = "traces/gpt4omini_vision/rollouts.db"
+output_dir = "traces/gpt4omini_vision/sft"
 
 # Quality filters
 min_steps_per_episode = 5
@@ -205,7 +197,7 @@ uvx synth-ai filter \
 
 **Expected output:**
 ```
-📂 Loading traces from traces/gpt5nano_vision/rollouts.db
+📂 Loading traces from traces/gpt4omini_vision/rollouts.db
    Total episodes: 100
    Total steps: 4,923
 
@@ -222,8 +214,8 @@ uvx synth-ai filter \
    ✓ Final dataset: 4,190 samples
 
 ✂️ Splitting train/val (90%/10%)...
-   ✓ Train: 3,771 samples → traces/gpt5nano_vision/sft/train.jsonl
-   ✓ Val: 419 samples → traces/gpt5nano_vision/sft/val.jsonl
+   ✓ Train: 3,771 samples → traces/gpt4omini_vision/sft/train.jsonl
+   ✓ Val: 419 samples → traces/gpt4omini_vision/sft/val.jsonl
 
 ✅ Filter complete!
 ```
@@ -236,7 +228,7 @@ Check the SFT JSONL format:
 
 ```bash
 # Inspect first sample
-head -1 traces/gpt5nano_vision/sft/train.jsonl | jq .
+head -1 traces/gpt4omini_vision/sft/train.jsonl | jq .
 ```
 
 **Expected format:**
@@ -282,7 +274,7 @@ head -1 traces/gpt5nano_vision/sft/train.jsonl | jq .
     "step": 12,
     "seed": 42,
     "has_image": true,
-    "model": "gpt-5-nano"
+    "model": "gpt-4o-mini-2024-07-18"
   }
 }
 ```
@@ -301,8 +293,8 @@ export BACKEND_BASE_URL="https://synth-backend-dev-docker.onrender.com/api"
 uvx synth-ai train \
   --type sft \
   --config configs/vision_sft/crafter_qwen3vl_8b_gpt5nano.toml \
-  --dataset traces/gpt5nano_vision/sft/train.jsonl \
-  --eval-dataset traces/gpt5nano_vision/sft/val.jsonl \
+  --dataset traces/gpt4omini_vision/sft/train.jsonl \
+  --eval-dataset traces/gpt4omini_vision/sft/val.jsonl \
   --env-file backend/.env.dev
 ```
 
@@ -313,15 +305,16 @@ uvx synth-ai train \
 ```bash
 # Terminal 1: Serve task app
 cd /Users/joshpurtell/Documents/GitHub/synth-ai
-uvx synth-ai serve \
-  --task-app examples/task_apps/crafter/task_app/synth_envs_hosted/main.py \
-  --port 8000
+uvx synth-ai deploy grpo-crafter-task-app \
+  --runtime uvicorn \
+  --port 8000 \
+  --trace traces/v3
 
 # Terminal 2: Collect traces
 export OPENAI_API_KEY="sk-..."
 uvx synth-ai eval \
   --config examples/qwen_vl/configs/eval_gpt5nano_vision.toml \
-  --output-dir traces/gpt5nano_vision
+  --trace-db traces/gpt4omini_vision/rollouts.db
 
 # Terminal 2: Filter and export
 uvx synth-ai filter \
@@ -333,8 +326,8 @@ export BACKEND_BASE_URL="https://synth-backend-dev-docker.onrender.com/api"
 uvx synth-ai train \
   --type sft \
   --config configs/vision_sft/crafter_qwen3vl_8b_gpt5nano.toml \
-  --dataset /Users/joshpurtell/Documents/GitHub/synth-ai/traces/gpt5nano_vision/sft/train.jsonl \
-  --eval-dataset /Users/joshpurtell/Documents/GitHub/synth-ai/traces/gpt5nano_vision/sft/val.jsonl \
+  --dataset /Users/joshpurtell/Documents/GitHub/synth-ai/traces/gpt4omini_vision/sft/train.jsonl \
+  --eval-dataset /Users/joshpurtell/Documents/GitHub/synth-ai/traces/gpt4omini_vision/sft/val.jsonl \
   --env-file backend/.env.dev
 ```
 
@@ -345,7 +338,7 @@ uvx synth-ai train \
 | Step | Duration | Cost | Notes |
 |------|----------|------|-------|
 | 1. Serve | Continuous | Free | Local or Modal |
-| 2. Eval (100 episodes) | 30-60 min | ~$1-2 | OpenAI gpt-5-nano |
+| 2. Eval (100 episodes) | 30-60 min | ~$1-2 | OpenAI gpt-4o-mini |
 | 3. Filter | < 5 min | Free | Local processing |
 | 4. SFT (2 epochs) | 2-4 hrs | ~$21 | 2x H200 on Modal |
 
@@ -364,12 +357,12 @@ uvx synth-ai eval --config configs/eval_gpt5nano_vision.toml
 # Collect from gpt-4o-mini (stronger teacher)
 uvx synth-ai eval --config configs/eval_gpt4o_mini_vision.toml
 
-# Collect from Qwen2-VL (for comparison)
-uvx synth-ai eval --config configs/eval_qwen2vl_vision.toml
+# Collect from Qwen3-VL (for comparison)
+uvx synth-ai eval --config configs/eval_qwen3vl_vision.toml
 
 # Merge and filter all traces
 uvx synth-ai filter \
-  --input-dbs traces/gpt5nano_vision/rollouts.db,traces/gpt4o_mini_vision/rollouts.db \
+  --input-dbs traces/gpt4omini_vision/rollouts.db,traces/qwen3vl_vision/rollouts.db \
   --output-dir traces/merged_vision/sft \
   --config configs/filter_vision_sft.toml
 ```
@@ -402,7 +395,7 @@ curl http://localhost:8000/health
 ```
 
 ### Traces not saving
-Ensure `collect_traces = true` in eval config and `trace_db` path is writable.
+Ensure you pass `--trace-db` (or accept the default) so traces land in a SQLite/Turso database.
 
 ### Filter removes all samples
 Lower quality thresholds:
@@ -420,4 +413,3 @@ min_achievements_per_episode = 0  # Allow episodes with no achievements
 - **Eval Config Schema:** `synth-ai eval --help`
 - **Filter Config Schema:** `synth-ai filter --help`
 - **Full Pipeline:** See `/Users/joshpurtell/Documents/GitHub/monorepo/vision_sft_rl.txt`
-

examples/qwen_vl/collect_vision_traces.py

@@ -22,13 +22,13 @@ Usage:
       --max-steps 50 \
       --output-dir traces/gpt5nano_vision
 
-  # Collect with Qwen2-VL via synth
+  # Collect with Qwen3-VL via synth
   uv run python examples/qwen_vl/collect_vision_traces.py \
-      --model Qwen/Qwen2-VL-7B-Instruct \
+      --model Qwen/Qwen3-VL-8B-Instruct \
       --provider synth \
       --episodes 100 \
       --max-steps 50 \
-      --output-dir traces/qwen2vl_vision
+      --output-dir traces/qwen3vl_vision
 """
 
 from __future__ import annotations
@@ -333,7 +333,7 @@ async def main() -> None:
     parser.add_argument(
         "--model",
         required=True,
-        help="Model name (e.g., gpt-5-nano, Qwen/Qwen2-VL-7B-Instruct)",
+        help="Model name (e.g., gpt-5-nano, Qwen/Qwen3-VL-8B-Instruct)",
     )
     parser.add_argument(
         "--provider",

examples/qwen_vl/configs/crafter_rl_vision_qwen3vl4b.toml

@@ -1,19 +1,9 @@
-# Crafter RL with Vision - Qwen3-VL-4B
-#
-# This configuration runs online RL (GRPO/GSPO) with a vision-language model
-# using the same Crafter task app that generates image observations for SFT data.
-#
-# Model: Qwen/Qwen3-VL-4B (smaller, faster for testing)
-# Task App: grpo-crafter-task-app (Modal deployed, supports vision)
-# Policy: crafter-react with use_vision=true, image_only_mode=true
-
 [algorithm]
 type = "online"
 method = "policy_gradient"
 variety = "gspo"
 
 [services]
-# Replace with the Modal URL printed by `uvx synth-ai modal-serve grpo-crafter`
 task_url = "https://YOUR-MODAL-TASK-APP.modal.run"
 
 [compute]
@@ -30,8 +20,6 @@ tensor_parallel = 1
 [vllm]
 tensor_parallel_size = 1
 max_model_len = 4096
-# Vision-specific settings
-limit_mm_per_prompt = { "image": 1 }  # Max 1 image per prompt
 
 [reference]
 placement = "none"
@@ -40,88 +28,83 @@ placement = "none"
 base = "Qwen/Qwen3-VL-4B-Instruct"
 trainer_mode = "lora"
 label = "crafter-rl-vision-qwen3vl4b"
-supports_vision = true  # Enable vision support
+supports_vision = true
 
 [lora]
 r = 16
 alpha = 32
 dropout = 0.05
-target_modules = ["all-linear"]
-# Note: will automatically include mm_projector for vision models
+target_modules = [ "all-linear",]
 
 [rollout]
 env_name = "crafter"
-max_turns = 10  # 10 steps per episode for faster testing
+max_turns = 10
 episodes_per_batch = 2
 policy_name = "crafter-react"
-max_concurrent_rollouts = 4  # Lower for vision models (memory)
+max_concurrent_rollouts = 4
 batches_per_step = 2
-ops = ["agent", "env"]
-
-  [rollout.env_config]
-  difficulty = "easy"
-  
-    [rollout.env_config.step_rewards]
-    enabled = true
-    mode = "decision_stepwise"
-    strategy = "consistent"
-    indicator_lambda = 1.0
-    step_beta = 0.0
-
-  [rollout.policy_config]
-  # Vision-specific policy settings
-  use_vision = true  # Enable vision input
-  image_only_mode = true  # Use only images, no text observations
-  temperature = 0.6  # Slightly higher for exploration
-  top_p = 0.95
-  max_tokens = 512
-  max_llm_calls = 10
+ops = [ "agent", "env",]
 
 [evaluation]
-instances = 8  # Lower for faster vision evals
+instances = 8
 every_n_iters = 5
-seeds = [0, 1, 2, 3, 4, 5, 6, 7]
+seeds = [ 0, 1, 2, 3, 4, 5, 6, 7,]
 
 [training]
 num_epochs = 1
-iterations_per_epoch = 3  # Shorter for integration test
+iterations_per_epoch = 3
 gradient_accumulation_steps = 2
 max_accumulated_minibatch = 1
 max_turns = 10
-batch_size = 2  # Smaller for vision models
+batch_size = 2
 group_size = 2
 learning_rate = 5e-5
 log_interval = 1
 weight_sync_interval = 1
 event_rewards_kind = "unique"
-async_semaphore_max = 2  # Lower concurrency for vision
-
-# Enable dense decision rewards
+async_semaphore_max = 2
 step_rewards_enabled = true
 step_rewards_mode = "decision_stepwise"
 step_rewards_indicator_lambda = 1.0
 step_rewards_beta = 0.0
 step_rewards_strategy = "consistent"
+max_images_per_message = 1
+supports_vision = true
+
+[tags]
+experiment = "crafter_rl_vision_qwen3vl4b"
+task = "crafter_agent_vision"
+model_size = "4b"
+vision_enabled = true
+image_only = true
 
-# Vision-specific training settings
-max_images_per_message = 1  # Limit images for memory
-supports_vision = true  # Enable vision training path
+[vllm.limit_mm_per_prompt]
+image = 1
+
+[rollout.env_config]
+difficulty = "easy"
+
+[rollout.policy_config]
+use_vision = true
+image_only_mode = true
+temperature = 0.6
+top_p = 0.95
+max_tokens = 512
+max_llm_calls = 10
 
 [training.weight_sync]
 enable = true
-targets = ["policy"]
+targets = [ "policy",]
 mode = "direct"
 direct = true
 verify_every_k = 0
 
-[judge]
-type = "env"  # Use environment rewards only (simpler for testing)
+[judge.options]
 timeout_s = 30
 
-[tags]
-experiment = "crafter_rl_vision_qwen3vl4b"
-task = "crafter_agent_vision"
-model_size = "4b"
-vision_enabled = true
-image_only = true
-
+[rollout.env_config.step_rewards]
+enabled = true
+mode = "decision_stepwise"
+strategy = "consistent"
+indicator_lambda = 1.0
+step_beta = 0.0

examples/qwen_vl/configs/crafter_vlm_sft_example.toml

@@ -7,7 +7,7 @@ method = "sft"
 variety = "lora"
 
 [job]
-model = "Qwen/Qwen2-VL-7B-Instruct"  # or Qwen/Qwen3-VL-8B
+model = "Qwen/Qwen3-VL-8B-Instruct"  # or Qwen/Qwen3-VL-4B-Instruct
 # Dataset from collect_vision_traces.py → export_to_sft.py
 data = "traces/gpt5nano_vision/train.jsonl"
 
@@ -57,4 +57,3 @@ task = "crafter"
 modality = "vision"
 data_source = "collected_traces"
 model_family = "qwen_vl"
-