synth-ai 0.2.9.dev3__py3-none-any.whl → 0.2.9.dev4__py3-none-any.whl

examples/finetuning_old/_backup_synth_qwen/example_log.md

@@ -0,0 +1,324 @@
+### Quickstart (using config.toml)
+
+The defaults come from `examples/finetuning/synth_qwen/config.toml`. Override with env vars only when needed.
+
+#### 1) Rollouts → v3 traces
+```bash
+set -a; source .env 2>/dev/null || true; set +a
+uvpm examples.finetuning.synth_qwen.run_crafter_qwen4b
+```
+
+Notes:
+- Model, episodes, steps, difficulty, temperature, tool choice, etc. are taken from `config.toml`.
+- Only API key and v3 db path are specified here.
+
+Example output (abridged):
+```text
+✅ Crafter service is healthy
+🚀 Running 10 episodes (concurrency=5)...
+✅ Completed 10 episodes in ~366s
+📊 EVALUATION RESULTS
+Episodes completed: 10/10
+Average reward per episode: 1.10
+Average steps per episode: 87.00
+💾 Results: traces/v3/synth_ai.db
+```
+
+#### 2) Filter traces → SFT JSONL
+```bash
+uvpm examples.finetuning.synth_qwen.filter_traces_achievements
+```
+
+Notes:
+- Filter settings (achievements, thresholds, output path) are defined in `config.toml`.
+
+Example output:
+```text
+Using database: sqlite+aiosqlite:///$PWD/traces/v3/synth_ai.db/dbs/default/data
+Output file: ft_data/qwen4b_crafter_sft_collect_wood.jsonl
+✅ Wrote 13 examples from 13 sessions
+```
+
+#### 3) Kick off SFT (prod)
+```bash
+set -a; source .env 2>/dev/null || true; set +a
+uvpm examples.finetuning.synth_qwen.sft_kickoff
+```
+
+Notes:
+- Base model and training JSONL path come from `config.toml`.
+
+Example output (abridged):
+```text
+🚀 Starting Qwen 4B SFT
+⏳ poll ...
+🟢 Qwen4B SFT fine-tune succeeded → ft:Qwen/Qwen3-4B-Instruct-2507:ftjob-22
+```
+
+#### 4) Rollouts with fine-tuned model
+```bash
+set -a; source .env 2>/dev/null || true; set +a
+CRAFTER_MODEL="ft:Qwen/Qwen3-4B-Instruct-2507:ftjob-22" uvpm examples.finetuning.synth_qwen.run_crafter_qwen4b
+```
+
+Notes:
+- Replace `ftjob-22` with the job id printed by your SFT step.
+- If you see 401 Invalid API key, export the prod key: `export SYNTH_API_KEY="$SYNTH_API_KEY_PROD"`.
+
+Example output (abridged):
+```text
+✅ Model warmed up successfully!
+🚀 Running 10 episodes (concurrency=5)...
+✅ Completed 10 episodes in ~480s
+📊 EVALUATION RESULTS
+Episodes completed: 10/10
+Average reward per episode: 1.60
+Average steps per episode: 90.80
+Achievements: collect_wood in 6/10 episodes
+💾 Results: traces/v3/synth_ai.db
+```
+
+
+
+```
+joshuapurtell@Mac synth-ai % bash examples/finetuning/synth_qwen/run_demo.sh
+Synth Qwen4B finetuning demo (Crafter)
+
+Run rollouts to generate v3 traces now? [Y/n]: Y
+Using config defaults from examples/finetuning/synth_qwen/config.toml (override below if desired).
+Model id [Enter=use config]: 
+Episodes [Enter=use config]: 5
+Max steps [Enter=use config]: 5
+Difficulty [Enter=use config]: 
+Enable think mode? (1/0) [Enter=0]: 
+
+Running rollouts (v3 tracing)...
+Detected SYNTH_API_KEY (sk_liv...ac95). Use this key? [Y/n]: N
+Use SYNTH_API_KEY_PROD (sk_liv...a2a4)? [y/N]: y
+[PATCH] Attempting to apply Crafter deterministic patch...
+[PATCH] Patching crafter.Env._balance_object...
+[PATCH] crafter.Env._balance_object patched.
+[PATCH] Attempting to apply Crafter serialization patch v3...
+[PATCH] Adding enhanced save/load methods to crafter.Env...
+[PATCH] crafter.Env.save() and load() methods added (v3).
+[PATCH] Crafter serialization patch v3 complete.
+[PATCH] Attempting to apply simplified Crafter world configuration patch...
+[PATCH] Simplified Crafter world configuration patch complete.
+[PATCH] Available configs: easy, normal, hard, peaceful
+🔧 Using Synth base URL = https://agent-learning.onrender.com/api
+🔇 Quiet mode enabled - suppressing verbose logs
+✅ Crafter service is healthy: {'status': 'ok', 'supported_environments': ['CrafterClassic', 'CrafterCustom']}
+
+🔥 Warming up Qwen/Qwen3-4B-Instruct-2507 on Synth backend...
+✅ Warmed Qwen/Qwen3-4B-Instruct-2507 in 3s        arming elapsed=0s
+✅ Model warmed up successfully!
+
+🚀 Starting sqld daemon for v3 tracing...
+✅ sqld daemon started
+
+📊 V3 Tracing enabled. Traces will be saved to: traces/v3/synth_ai.db
+   Experiment: crafter_lm_synth_Qwen/Qwen3-4B-Instruct-2507_20250808_165304
+
+🚀 Running 5 episodes (concurrency=5)...
+
+📤 Starting episodes...
+Episode 1: 100%|███████████████| 5/5 [00:18<00:00,  3.71s/it, tc=1, act=3, tok=78, in=860, tps=22.4]
+Episode 0: 100%|██████████████| 5/5 [00:19<00:00,  3.89s/it, tc=1, act=3, tok=94, in=861, tps=22.79]
+Episode 2: 100%|██████████████| 5/5 [00:18<00:00,  3.74s/it, tc=1, act=2, tok=73, in=832, tps=21.06]
+Episode 4: 100%|██████████████| 5/5 [00:19<00:00,  3.90s/it, tc=1, act=3, tok=91, in=818, tps=17.34]
+Episode 3: 100%|██████████████| 5/5 [00:20<00:00,  4.15s/it, tc=1, act=2, tok=86, in=859, tps=21.96]
+
+✅ Completed 5 episodes in 23.57 seconds
+
+==================================================
+📊 EVALUATION RESULTS
+==================================================
+Episodes completed: 5/5
+Failed episodes: 0
+Total reward: 1.00
+Average reward per episode: 0.20
+Total steps: 68
+Average steps per episode: 13.60
+
+Seeds used:
+  Episode 0: seed 1
+  Episode 1: seed 2
+  Episode 2: seed 3
+  Episode 3: seed 4
+  Episode 4: seed 5
+Unique achievements unlocked: 1
+
+Achievements unlocked:
+  - collect_sapling: 1 episodes (20.0%)
+
+Action counts (total: 68):
+  - move_right: 25 (36.8%)
+  - do: 19 (27.9%)
+  - move_down: 10 (14.7%)
+  - move_left: 5 (7.4%)
+  - place_stone: 3 (4.4%)
+  - move_up: 3 (4.4%)
+  - make_wood_sword: 1 (1.5%)
+  - make_stone_pickaxe: 1 (1.5%)
+  - make_wood_pickaxe: 1 (1.5%)
+
+💾 Results available in Turso database: traces/v3/synth_ai.db
+   Experiment ID: exp_a0b091fd12c6
+   Use the filter_traces_sft_turso.py script to extract fine-tuning data
+
+Markdown row:
+| Qwen/Qwen3-4B-Instruct-2507 |       5 |     0.20 |         0.020 |        0.001 |       68.000 |   13.600 |
+
+✅ Stopped sqld daemon
+
+Filter v3 traces into SFT JSONL now? [Y/n]: Y
+Using DB: sqlite+aiosqlite:////Users/joshuapurtell/Documents/GitHub/synth-ai/traces/v3/synth_ai.db/dbs/default/data
+You can override filter options; Enter to use config defaults.
+Required achievements (space-separated) [Enter=config]: 
+Restrict to models (space-separated) [Enter=all]: 
+Output JSONL path [Enter=config]: 
+Min total reward [Enter=config]: 1
+Max total cost [Enter=config]: 
+Max total tokens [Enter=config]: 
+
+Filtering traces to SFT JSONL...
+[PATCH] Attempting to apply Crafter deterministic patch...
+[PATCH] Patching crafter.Env._balance_object...
+[PATCH] crafter.Env._balance_object patched.
+[PATCH] Attempting to apply Crafter serialization patch v3...
+[PATCH] Adding enhanced save/load methods to crafter.Env...
+[PATCH] crafter.Env.save() and load() methods added (v3).
+[PATCH] Crafter serialization patch v3 complete.
+[PATCH] Attempting to apply simplified Crafter world configuration patch...
+[PATCH] Simplified Crafter world configuration patch complete.
+[PATCH] Available configs: easy, normal, hard, peaceful
+🤖 Modal/Synth FT Filter (achievements)
+Using database: sqlite+aiosqlite:////Users/joshuapurtell/Documents/GitHub/synth-ai/traces/v3/synth_ai.db/dbs/default/data
+Output file: ft_data/qwen4b_crafter_sft.jsonl
+Filters: {
+  "required_achievements": [
+    "collect_wood"
+  ],
+  "models": [
+    "Qwen/Qwen3-4B-Instruct-2507"
+  ],
+  "min_total_reward": 1.0,
+  "max_cost": 10.0,
+  "max_tokens": 100000
+}
+
+✅ Wrote 23 examples from 23 sessions
+
+Kick off SFT training job now? [Y/n]: Y
+Enter overrides for training job; Enter to use config.
+Base model [Enter=config]: 
+Training JSONL path [Enter=config]: 
+
+Starting SFT job...
+Detected SYNTH_API_KEY (sk_liv...a2a4). Use this key? [Y/n]: n
+Use SYNTH_API_KEY_PROD (sk_liv...a2a4)? [y/N]: Y
+🚀 Starting Qwen 4B SFT
+⏳ poll 1/20 – status = queued
+⏳ poll 2/20 – status = queued
+⏳ poll 3/20 – status = queued
+⏳ poll 4/20 – status = queued
+⏳ poll 5/20 – status = queued
+⏳ poll 6/20 – status = queued
+⏳ poll 7/20 – status = queued
+⏳ poll 8/20 – status = queued
+⏳ poll 9/20 – status = running
+⏳ poll 10/20 – status = queued
+⏳ poll 11/20 – status = queued
+⏳ poll 12/20 – status = succeeded
+🟢 Qwen4B SFT fine-tune succeeded → ft:Qwen/Qwen3-4B-Instruct-2507:ftjob-6cedf721e0ca4c80968834b71e2bdace
+⏱️ wall-clock: 249.3s | trained_tokens: 41777
+Captured fine-tuned model id: ft:Qwen/Qwen3-4B-Instruct-2507:ftjob-6cedf721e0ca4c80968834b71e2bdace
+SFT logs saved to: logs/sft_kickoff_20250808_165436.log
+
+Roll out fine-tuned model 'ft:Qwen/Qwen3-4B-Instruct-2507:ftjob-6cedf721e0ca4c80968834b71e2bdace' in Crafter now? [y/N]: Y
+Episodes [Enter=config]: 5
+Max steps [Enter=config]: 5
+Difficulty [Enter=config]: 
+Enable think mode? (1/0) [Enter=0]: 
+
+Running rollouts with fine-tuned model...
+[PATCH] Attempting to apply Crafter deterministic patch...
+[PATCH] Patching crafter.Env._balance_object...
+[PATCH] crafter.Env._balance_object patched.
+[PATCH] Attempting to apply Crafter serialization patch v3...
+[PATCH] Adding enhanced save/load methods to crafter.Env...
+[PATCH] crafter.Env.save() and load() methods added (v3).
+[PATCH] Crafter serialization patch v3 complete.
+[PATCH] Attempting to apply simplified Crafter world configuration patch...
+[PATCH] Simplified Crafter world configuration patch complete.
+[PATCH] Available configs: easy, normal, hard, peaceful
+🔧 Using Synth base URL = https://agent-learning.onrender.com/api
+🔇 Quiet mode enabled - suppressing verbose logs
+✅ Crafter service is healthy: {'status': 'ok', 'supported_environments': ['CrafterClassic', 'CrafterCustom']}
+
+🔥 Warming up ft:Qwen/Qwen3-4B-Instruct-2507:ftjob-6cedf721e0ca4c80968834b71e2bdace on Synth backend...
+⏳ Warming ft:Qwen/Qwen3-4B-Instruct-2507:ftjob-6cedf721e0ca4c80968834b71e2bdace [|] status=timeout elapsed=10⏳ Warming ft:Qwen/Qwen3-4B-Instruct-2507:ftjob-6cedf721e0ca4c80968834b71e2bdace [/] status=timeout elapsed=21✅ Warmed ft:Qwen/Qwen3-4B-Instruct-2507:ftjob-6cedf721e0ca4c80968834b71e2bdace in 22s        
+✅ Model warmed up successfully!
+
+🚀 Starting sqld daemon for v3 tracing...
+✅ sqld daemon started
+
+📊 V3 Tracing enabled. Traces will be saved to: traces/v3/synth_ai.db
+   Experiment: crafter_lm_synth_ft:Qwen/Qwen3-4B-Instruct-2507:ftjob-6cedf721e0ca4c80968834b71e2bdace_20250808_165943
+
+🚀 Running 5 episodes (concurrency=5)...
+
+📤 Starting episodes...
+Episode 2: 100%|██████████████| 5/5 [00:48<00:00,  9.80s/it, tc=1, act=3, tok=75, in=833, tps=15.24]
+Episode 1: 100%|██████████████| 5/5 [00:52<00:00, 10.48s/it, tc=1, act=3, tok=73, in=840, tps=10.74]
+Episode 3: 100%|██████████████| 5/5 [00:54<00:00, 10.88s/it, tc=1, act=3, tok=79, in=834, tps=15.59]
+Episode 4: 100%|██████████████| 5/5 [00:54<00:00, 10.90s/it, tc=1, act=3, tok=75, in=817, tps=11.93]
+Episode 0: 100%|███████████████| 5/5 [00:56<00:00, 11.38s/it, tc=1, act=2, tok=91, in=850, tps=5.53]
+
+✅ Completed 5 episodes in 58.29 seconds
+
+==================================================
+📊 EVALUATION RESULTS
+==================================================
+Episodes completed: 5/5
+Failed episodes: 0
+Total reward: 3.00
+Average reward per episode: 0.60
+Total steps: 72
+Average steps per episode: 14.40
+
+Seeds used:
+  Episode 0: seed 1
+  Episode 1: seed 2
+  Episode 2: seed 3
+  Episode 3: seed 4
+  Episode 4: seed 5
+Unique achievements unlocked: 2
+
+Achievements unlocked:
+  - collect_sapling: 2 episodes (40.0%)
+  - collect_wood: 1 episodes (20.0%)
+
+Action counts (total: 72):
+  - move_right: 25 (34.7%)
+  - do: 19 (26.4%)
+  - move_down: 8 (11.1%)
+  - move_left: 8 (11.1%)
+  - move_up: 6 (8.3%)
+  - place_stone: 2 (2.8%)
+  - place_table: 2 (2.8%)
+  - make_wood_pickaxe: 1 (1.4%)
+  - place_plant: 1 (1.4%)
+
+💾 Results available in Turso database: traces/v3/synth_ai.db
+   Experiment ID: exp_56d8e29cbf5a
+   Use the filter_traces_sft_turso.py script to extract fine-tuning data
+
+Markdown row:
+| ft:Qwen/Qwen3-4B-Instruct-2507:ftjob-6cedf721e0ca4c80968834b71e2bdace |       5 |     0.60 |         0.240 |        0.012 |       72.000 |   14.400 |
+
+✅ Stopped sqld daemon
+
+Done. You can re-run this script to repeat steps as needed.
+joshuapurtell@Mac synth-ai % 
+```

examples/finetuning_old/_backup_synth_qwen/filter_traces.py

@@ -0,0 +1,60 @@
+#!/usr/bin/env python3
+"""
+Filter v3 Crafter traces into an SFT-ready JSONL using the maintained
+Modal/Synth filter logic (no CLI needed). Intended to be run after
+collecting trajectories with the Crafter runner.
+
+Environment:
+- CRAFTER_DB_URL (default: sqlite:///traces_v3_lm_synth/traces.db)
+- OUTPUT_JSONL (default: ft_data/qwen4b_crafter_sft.jsonl)
+- MIN_TOTAL_REWARD (float, default: 1.0)
+- MIN_ACHIEVEMENTS (int, default: 0)
+- MAX_COST (float, default: 10.0)
+- MAX_TOKENS (int, default: 100000)
+- MODELS (optional, space-separated model names; default empty = all)
+"""
+
+import asyncio
+import json
+import os
+from typing import Any
+
+# Reuse the existing filtering implementation
+from synth_ai.environments.examples.crafter_classic.agent_demos.crafter_modal_ft.filter_traces_sft_turso import (
+    filter_traces_from_turso,
+)
+
+
+def build_config() -> dict[str, Any]:
+    models_env = os.getenv("MODELS", "").strip()
+    models: list[str] = models_env.split() if models_env else []
+    return {
+        "mode": "trajectory",
+        "filters": {
+            "min_total_reward": float(os.getenv("MIN_TOTAL_REWARD", "1.0")),
+            "min_achievements": int(os.getenv("MIN_ACHIEVEMENTS", "0")),
+            "max_cost": float(os.getenv("MAX_COST", "10.0")),
+            "max_tokens": int(os.getenv("MAX_TOKENS", "100000")),
+            "models": models,
+        },
+    }
+
+
+async def main() -> None:
+    db_url = os.getenv("CRAFTER_DB_URL", "sqlite:///traces_v3_lm_synth/traces.db")
+    output_path = os.getenv("OUTPUT_JSONL", "ft_data/qwen4b_crafter_sft.jsonl")
+    config = build_config()
+
+    print("🤖 Modal/Synth Fine-Tuning Data Filter (v3)")
+    print("Using database:", db_url)
+    print("Output file:", output_path)
+    print("Config:", json.dumps(config, indent=2))
+
+    num_examples, stats = await filter_traces_from_turso(db_url, output_path, config)
+
+    print("\n✅ Wrote", num_examples, "training examples to", output_path)
+    print("📊 Stats keys:", list(stats.keys()))
+
+
+if __name__ == "__main__":
+    asyncio.run(main())

examples/finetuning_old/_backup_synth_qwen/filter_traces_achievements.py

@@ -0,0 +1,239 @@
+#!/usr/bin/env python3
+"""
+Filter v3 Crafter traces into SFT JSONL requiring specific achievements.
+
+Environment:
+- CRAFTER_DB_URL (default: sqlite:///traces_v3_lm_synth/traces.db)
+- OUTPUT_JSONL (default: ft_data/qwen4b_crafter_sft_ach.jsonl)
+- REQUIRED_ACHIEVEMENTS (space-separated, default: collect_wood)
+- MIN_TOTAL_REWARD (float, default: 0.0)
+- MAX_COST (float, default: inf)
+- MAX_TOKENS (int, default: inf)
+- MODELS (optional, space-separated model names; default empty = all)
+- WINDOW_MODE=1 to emit per-turn user→assistant examples
+"""
+
+import asyncio
+import json
+import math
+import os
+import tomllib
+from pathlib import Path
+from typing import Any
+
+try:
+    # Preferred path (modal-specific)
+    from synth_ai.environments.examples.crafter_classic.agent_demos.crafter_modal_ft.filter_traces_sft_turso import (  # type: ignore
+        FinetuningDataExtractorV3,
+    )
+except Exception:  # pragma: no cover
+    try:
+        # Fallback path used in some dist builds
+        from synth_ai.environments.examples.crafter_classic.agent_demos.crafter_openai_ft.filter_traces_sft_turso import (  # type: ignore
+            FinetuningDataExtractorV3,
+        )
+    except Exception as _import_err:  # pragma: no cover
+        raise ImportError(
+            "Could not import FinetuningDataExtractorV3 from synth_ai.") from _import_err
+
+
+def env_list(name: str) -> list[str]:
+    val = os.getenv(name, "").strip()
+    return val.split() if val else []
+
+
+def normalize_db_url(raw: str) -> str:
+    # Accept file path or sqlite URLs; ensure async driver prefix
+    if raw.endswith(".db") and not raw.startswith("sqlite"):
+        return f"sqlite+aiosqlite:///{raw}"
+    if raw.startswith("sqlite+aiosqlite///"):
+        return raw
+    if raw.startswith("sqlite///") and raw.endswith(".db"):
+        return raw.replace("sqlite///", "sqlite+aiosqlite///")
+    return raw
+
+
+def build_filters() -> dict[str, Any]:
+    cfg_default = Path(__file__).with_name("config.toml")
+    cfg_path = os.getenv("CRAFTER_CONFIG", str(cfg_default))
+    cfg: dict[str, Any] = {}
+    if os.path.exists(cfg_path):
+        with open(cfg_path, "rb") as f:
+            cfg = tomllib.load(f)
+    fcfg = cfg.get("filter", {})
+    # Default: no required achievements gating unless provided via env/config
+    req = set(env_list("REQUIRED_ACHIEVEMENTS") or fcfg.get("required_achievements", []))
+    models = env_list("MODELS")
+    # Default: allow zero reward unless overridden
+    min_reward = float(os.getenv("MIN_TOTAL_REWARD", str(fcfg.get("min_total_reward", 0.0))))
+    max_cost = float(os.getenv("MAX_COST", str(fcfg.get("max_cost", math.inf))))
+    max_tokens = int(os.getenv("MAX_TOKENS", str(fcfg.get("max_tokens", 1_000_000_000))))
+    return {
+        "required_achievements": req,
+        "models": models,
+        "min_total_reward": min_reward,
+        "max_cost": max_cost,
+        "max_tokens": max_tokens,
+    }
+
+
+async def main() -> None:
+    cfg_default = Path(__file__).with_name("config.toml")
+    cfg_path = os.getenv("CRAFTER_CONFIG", str(cfg_default))
+    cfg: dict[str, Any] = {}
+    if os.path.exists(cfg_path):
+        with open(cfg_path, "rb") as f:
+            cfg = tomllib.load(f)
+    fcfg = cfg.get("filter", {})
+    tcfg = cfg.get("traces", {})
+
+    # Prefer env; else derive from config or repo-local v3 path
+    raw_db_url = os.getenv("CRAFTER_DB_URL", "")
+    if not raw_db_url:
+        db_path = fcfg.get("db_path")
+        if not db_path and tcfg.get("sqld_db_path"):
+            # derive the internal data file path from the sqld dir
+            db_path = str(Path(tcfg["sqld_db_path"]) / "dbs" / "default" / "data")
+        if db_path:
+            raw_db_url = f"sqlite+aiosqlite:///{db_path}"
+        else:
+            # Try repo-local default: traces/v3/synth_ai.db/dbs/default/data
+            repo_root = Path(__file__).resolve().parents[3]
+            candidate = repo_root / "traces" / "v3" / "synth_ai.db" / "dbs" / "default" / "data"
+            raw_db_url = f"sqlite+aiosqlite:///{candidate}"
+    db_url = normalize_db_url(raw_db_url)
+    output_path = os.getenv(
+        "OUTPUT_JSONL", fcfg.get("output_jsonl", "ft_data/qwen4b_crafter_sft_ach.jsonl")
+    )
+    filters = build_filters()
+    # Default: require >1 achievements unless explicitly specified in config/env
+    # If caller set REQUIRED_ACHIEVEMENTS or provided config 'required_achievements', we won't override.
+    # Otherwise, enforce min achievements via MIN_ACHIEVEMENTS (default 2)
+    if not filters.get("required_achievements"):
+        try:
+            min_ach = int(os.getenv("MIN_ACHIEVEMENTS", str(fcfg.get("min_achievements", 2))))
+        except Exception:
+            min_ach = 2
+        filters["min_achievements"] = min_ach
+
+    window_mode = os.getenv("WINDOW_MODE", "0") == "1"
+
+    print("🤖 Modal/Synth FT Filter (achievements)")
+    print("Using database:", db_url)
+    print("Output file:", output_path)
+    print(
+        "Filters:",
+        json.dumps(
+            {k: (list(v) if isinstance(v, set) else v) for k, v in filters.items()}, indent=2
+        ),
+    )
+    print("Window mode:", window_mode)
+
+    # Print distributions (achievements and rewards) before filtering for visibility
+    try:
+        import numpy as _np
+        from collections import Counter as _Counter
+        async with FinetuningDataExtractorV3(db_url) as _ex:
+            _sessions = await _ex.get_all_sessions()
+            _ach_counts: _Counter[str] = _Counter()
+            _rewards: list[float] = []
+            for _, _row in _sessions.iterrows():
+                _sid = _row["session_id"]
+                _ach = await _ex.get_session_achievements(_sid) or []
+                for _a in _ach:
+                    _ach_counts[_a] += 1
+                _met = await _ex.get_session_metrics(_sid)
+                try:
+                    _rewards.append(float(_met.get("total_reward", 0.0) or 0.0))
+                except Exception:
+                    pass
+            print(f"\nTotal sessions: {len(_sessions)}")
+            if _ach_counts:
+                print("\nAchievements by session (count):")
+                for _name, _c in sorted(_ach_counts.items(), key=lambda x: (-x[1], x[0])):
+                    print(f"  {_name}: {_c}")
+            if _rewards:
+                _r = _np.array(_rewards, dtype=float)
+                print("\nReward stats:")
+                print(f"  min={_r.min():.2f} median={_np.median(_r):.2f} mean={_r.mean():.2f} max={_r.max():.2f}")
+    except Exception:
+        pass
+
+    required: set[str] = filters["required_achievements"]
+    models: list[str] = filters["models"]
+    min_reward: float = filters["min_total_reward"]
+    max_cost: float = filters["max_cost"]
+    max_tokens: int = filters["max_tokens"]
+
+    stats: dict[str, Any] = {
+        "total_sessions": 0,
+        "kept_sessions": 0,
+        "total_examples": 0,
+    }
+
+    async with FinetuningDataExtractorV3(db_url) as extractor:
+        all_sessions = await extractor.get_all_sessions()
+        stats["total_sessions"] = len(all_sessions)
+
+        kept: list[str] = []
+        for _, row in all_sessions.iterrows():
+            session_id = row["session_id"]
+            metrics = await extractor.get_session_metrics(session_id)
+
+            if metrics["total_reward"] < min_reward:
+                continue
+            if metrics["total_cost"] > max_cost:
+                continue
+            if metrics["total_tokens"] > max_tokens:
+                continue
+
+            if models:
+                model_query = """
+                    SELECT DISTINCT model_name
+                    FROM events
+                    WHERE session_id = :session_id
+                      AND event_type = 'cais'
+                      AND model_name IS NOT NULL
+                    """
+                model_df = await extractor.db_manager.query_traces(
+                    model_query, {"session_id": session_id}
+                )
+                session_models = model_df["model_name"].tolist() if not model_df.empty else []
+                if not any(m in models for m in session_models):
+                    continue
+
+            # Respect either explicit required achievements OR min_achievements fallback
+            min_ach = int(filters.get("min_achievements", 0))
+            if required or min_ach > 0:
+                achievements = await extractor.get_session_achievements(session_id)
+                if not achievements:
+                    continue
+                if required:
+                    if not (required & set(achievements)):
+                        continue
+                else:
+                    if len(achievements) < min_ach:
+                        continue
+
+            kept.append(session_id)
+
+        stats["kept_sessions"] = len(kept)
+
+        if window_mode:
+            training_data = await extractor.extract_openai_window_format(kept)
+        else:
+            training_data = await extractor.extract_openai_format(kept)
+
+        Path(output_path).parent.mkdir(parents=True, exist_ok=True)
+        with open(output_path, "w") as f:
+            for ex in training_data:
+                f.write(json.dumps(ex) + "\n")
+        stats["total_examples"] = len(training_data)
+
+    print(
+        "\n✅ Wrote", stats["total_examples"], "examples from", stats["kept_sessions"], "sessions"
+    )
+
+
+if __name__ == "__main__":
+    asyncio.run(main())