Spaces:
Sleeping
Sleeping
Add SFT and GRPO run commands to README
Browse files
README.md
CHANGED
|
@@ -1,321 +1,225 @@
|
|
| 1 |
-
---
|
| 2 |
-
title: CyberSecurity_OWASP Environment Server
|
| 3 |
-
emoji: 🛡️
|
| 4 |
-
colorFrom: blue
|
| 5 |
-
colorTo: gray
|
| 6 |
-
sdk: docker
|
| 7 |
-
pinned: false
|
| 8 |
-
app_port: 8000
|
| 9 |
-
base_path: /web
|
| 10 |
-
tags:
|
| 11 |
-
- openenv
|
| 12 |
-
- cybersecurity
|
| 13 |
-
- owasp
|
| 14 |
-
---
|
| 15 |
-
|
| 16 |
-
# CyberSecurity_OWASP
|
| 17 |
-
|
| 18 |
-
[Hugging Face Space](https://huggingface.co/spaces/Humanlearning/CyberSecurity_OWASP) | [Mini-blog](blog/blog.md)
|
| 19 |
-
|
| 20 |
-
`CyberSecurity_OWASP` is an OpenEnv-compliant reinforcement-learning environment for a single LLM agent that performs a defensive authorization-repair workflow:
|
| 21 |
-
|
| 22 |
-
```text
|
| 23 |
-
inspect generated app + policy -> discover authorization bug -> submit diagnosis -> patch code -> preserve intended behavior
|
| 24 |
-
```
|
| 25 |
-
|
| 26 |
-
The current implementation includes a functional closed-loop MVP scenario: an invoices FastAPI-style app with one injected OWASP A01 BOLA/IDOR defect, config-driven curriculum settings, cache-backed scenario reset, an ephemeral app sandbox, multi-layer deterministic verifier checks, anti-cheat safeguards, JSONL episode artifacts, and decomposed reward.
|
| 27 |
-
|
| 28 |
-
## Diagrams
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
 | [Mini-blog](blog/blog.md)
|
| 19 |
+
|
| 20 |
+
`CyberSecurity_OWASP` is an OpenEnv-compliant reinforcement-learning environment for a single LLM agent that performs a defensive authorization-repair workflow:
|
| 21 |
+
|
| 22 |
+
```text
|
| 23 |
+
inspect generated app + policy -> discover authorization bug -> submit diagnosis -> patch code -> preserve intended behavior
|
| 24 |
+
```
|
| 25 |
+
|
| 26 |
+
The current implementation includes a functional closed-loop MVP scenario: an invoices FastAPI-style app with one injected OWASP A01 BOLA/IDOR defect, config-driven curriculum settings, cache-backed scenario reset, an ephemeral app sandbox, multi-layer deterministic verifier checks, anti-cheat safeguards, JSONL episode artifacts, and decomposed reward.
|
| 27 |
+
|
| 28 |
+
## Diagrams
|
| 29 |
+
|
| 30 |
+
[Architecture diagram](assets/architecture_diagram.svg) | [RL training flow diagram](assets/env_rl_training_flow_diagram.svg)
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
Editable Mermaid sources are available in `assets/architecture_diagram.mmd` and `assets/env_rl_training_flow_diagram.mmd`.
|
| 37 |
+
|
| 38 |
+
## Quick Start
|
| 39 |
+
|
| 40 |
+
```bash
|
| 41 |
+
uv sync --extra dev
|
| 42 |
+
uv run --extra dev pytest
|
| 43 |
+
uv run python scripts/generate_scenario_cache.py --train-per-bucket 3 --validation-per-bucket 3 --heldout-per-bucket 3
|
| 44 |
+
uv run server --port 8000
|
| 45 |
+
```
|
| 46 |
+
|
| 47 |
+
Then connect with the OpenEnv client:
|
| 48 |
+
|
| 49 |
+
```python
|
| 50 |
+
from CyberSecurity_OWASP import CyberSecurityOWASPAction, CyberSecurityOWASPEnv
|
| 51 |
+
|
| 52 |
+
with CyberSecurityOWASPEnv(base_url="http://localhost:8000") as env:
|
| 53 |
+
result = env.reset(seed=7)
|
| 54 |
+
print(result.observation.task_brief)
|
| 55 |
+
result = env.step(CyberSecurityOWASPAction(tool_name="list_routes"))
|
| 56 |
+
print(result.observation.last_tool_result)
|
| 57 |
+
```
|
| 58 |
+
|
| 59 |
+
## Action Space
|
| 60 |
+
|
| 61 |
+
The agent emits one JSON action at a time:
|
| 62 |
+
|
| 63 |
+
```json
|
| 64 |
+
{"tool_name":"read_file","arguments":{"path":"app/routes/invoices.py"}}
|
| 65 |
+
```
|
| 66 |
+
|
| 67 |
+
Supported tools:
|
| 68 |
+
|
| 69 |
+
- `inspect_policy_graph`
|
| 70 |
+
- `list_routes`
|
| 71 |
+
- `read_openapi`
|
| 72 |
+
- `read_file`
|
| 73 |
+
- `search_code`
|
| 74 |
+
- `send_local_request`
|
| 75 |
+
- `compare_identities`
|
| 76 |
+
- `submit_diagnosis`
|
| 77 |
+
- `patch_file`
|
| 78 |
+
- `run_visible_tests`
|
| 79 |
+
- `submit_fix`
|
| 80 |
+
- `noop`
|
| 81 |
+
|
| 82 |
+
Tools are phase-gated:
|
| 83 |
+
|
| 84 |
+
- `discover`: inspect policy/routes/files, run safe local requests, compare identities, submit diagnosis.
|
| 85 |
+
- `patch`: read/search, patch editable app files, run visible tests, submit final fix.
|
| 86 |
+
- `done`: stable terminal observation only.
|
| 87 |
+
|
| 88 |
+
## Reward
|
| 89 |
+
|
| 90 |
+
Terminal reward uses stable components:
|
| 91 |
+
|
| 92 |
+
```python
|
| 93 |
+
{
|
| 94 |
+
"discovery": 0.0,
|
| 95 |
+
"security": 0.0,
|
| 96 |
+
"regression": 0.0,
|
| 97 |
+
"public_routes": 0.0,
|
| 98 |
+
"patch_quality": 0.0,
|
| 99 |
+
"visible_tests": 0.0,
|
| 100 |
+
"safety": 0.0,
|
| 101 |
+
"anti_cheat": 0.0,
|
| 102 |
+
"terminal_total": 0.0,
|
| 103 |
+
"progressive": 0.0,
|
| 104 |
+
"step_penalty": 0.0,
|
| 105 |
+
"speed_bonus": 0.0,
|
| 106 |
+
"token_penalty": 0.0,
|
| 107 |
+
"behavior_penalty": 0.0,
|
| 108 |
+
"train_total": 0.0,
|
| 109 |
+
"total": 0.0,
|
| 110 |
+
}
|
| 111 |
+
```
|
| 112 |
+
|
| 113 |
+
The verifier rewards blocking the hidden exploit while preserving legitimate owner/admin behavior and intentionally public routes. Terminal scoring requires visible checks, hidden authorization checks, a policy-oracle matrix, regression checks, public-route preservation, and patch-quality checks. It penalizes deny-all fixes, hardcoded IDs, repeated/invalid action patterns, hidden file probes, external URL attempts, and test/fixture tampering.
|
| 114 |
+
|
| 115 |
+
Training can enable dense rewards with `CYBERSECURITY_OWASP_REWARD_MODE=dense_train`.
|
| 116 |
+
Dense mode adds configurable progressive rewards, small efficiency penalties, and capped behavior penalties from `training/configs/grpo_small.yaml`; evaluation defaults to sparse terminal scoring.
|
| 117 |
+
|
| 118 |
+
## Scenario Cache And Generation
|
| 119 |
+
|
| 120 |
+
Scenario generation is an offline/cache-prep concern. `reset(seed)` asks the `CurriculumController` for a difficulty tier and target weakness, then loads a validated executable bundle from the scenario cache when `CYBERSECURITY_OWASP_SCENARIO_CACHE_MODE=require`. Local development defaults to `fallback`, which compiles deterministically on a cache miss.
|
| 121 |
+
|
| 122 |
+
The scenario/curriculum author is config-driven through `configs/scenario_authoring.small.json`. The default offline author model is `deepseek-ai/DeepSeek-V4-Pro` with Hugging Face provider settings, thinking mode enabled, `temperature=1.0`, and `top_p=1.0`. This model config is for scenario authoring, not the RL policy model.
|
| 123 |
+
|
| 124 |
+
The cache bundle contract is:
|
| 125 |
+
|
| 126 |
+
- `scenario.json`
|
| 127 |
+
- `app_source/`
|
| 128 |
+
- `policy_graph.json`
|
| 129 |
+
- `visible_tests.py`
|
| 130 |
+
- `hidden_tests.py`
|
| 131 |
+
- `oracle_tests.py`
|
| 132 |
+
- `expected_exploit_trace.json`
|
| 133 |
+
- `reward_config.json`
|
| 134 |
+
- `metadata.json`
|
| 135 |
+
|
| 136 |
+
Cache keys include difficulty, authorization bug type, app family, framework, policy shape, tenant model, exploit depth, patch scope, regression risk, generator version, verifier version, and scenario hash.
|
| 137 |
+
|
| 138 |
+
The MVP compiler currently generates:
|
| 139 |
+
|
| 140 |
+
- invoices domain policy graph;
|
| 141 |
+
- bounded adversarial target metadata such as same-role cross-object access, cross-tenant access, public-route overlocking traps, alternate route/service reachability, or visible-test-only edge cases;
|
| 142 |
+
- randomized users, tenants, invoices, and IDs;
|
| 143 |
+
- generated app files under `app/`;
|
| 144 |
+
- visible tests under `tests/test_visible.py`;
|
| 145 |
+
- hidden facts, oracle tuples, scenario family metadata, and verifier targets kept out of observations.
|
| 146 |
+
|
| 147 |
+
Additional domains and bug families are scaffolded for extension.
|
| 148 |
+
|
| 149 |
+
## Runtime Components
|
| 150 |
+
|
| 151 |
+
The OpenEnv runtime is split into small server modules:
|
| 152 |
+
|
| 153 |
+
- `server/curriculum.py` tracks mastery, weak spots, reward trend, and difficulty tier.
|
| 154 |
+
- `server/scenario_cache.py` writes and loads validated executable scenario bundles.
|
| 155 |
+
- `server/adversarial_designer.py` chooses safe synthetic scenario targets from tracked weaknesses.
|
| 156 |
+
- `server/scenario_factory.py` compiles the generated app during cache prep or local fallback.
|
| 157 |
+
- `server/app_sandbox.py` handles editable workspace reads, patches, local requests, and OpenAPI summaries.
|
| 158 |
+
- `server/action_tools.py` dispatches typed tools through the sandbox.
|
| 159 |
+
- `server/authz_oracle.py` builds the hidden allowed/denied user-resource-action matrix.
|
| 160 |
+
- `server/verifier.py` aggregates visible tests, hidden tests, oracle matrix, regression/public-route checks, and patch quality.
|
| 161 |
+
- `server/episode_logger.py` appends JSONL rollouts under `outputs/rollouts/`.
|
| 162 |
+
|
| 163 |
+
The agent sees partial observations only: product rules, fixture aliases, route summaries, visible test results, and action errors. Hidden tests, oracle tuples, injected bug labels, and held-out scenario-family labels stay internal.
|
| 164 |
+
|
| 165 |
+
## Testing
|
| 166 |
+
|
| 167 |
+
```bash
|
| 168 |
+
uv run --extra dev pytest
|
| 169 |
+
```
|
| 170 |
+
|
| 171 |
+
The suite covers model serialization, reset/step/state behavior, seed reproducibility, invalid actions, reward outcomes, anti-cheat checks, scripted rollout policies, curriculum selection, adversarial targeting, held-out scenario families, oracle checks, verifier aggregation, and episode artifact logging.
|
| 172 |
+
|
| 173 |
+
## Training Scaffold
|
| 174 |
+
|
| 175 |
+
Training files are under `training/`:
|
| 176 |
+
|
| 177 |
+
- `rollout.py`
|
| 178 |
+
- `reward_funcs.py`
|
| 179 |
+
- `train_grpo.py`
|
| 180 |
+
- `eval_before_after.py`
|
| 181 |
+
- `trackio_utils.py`
|
| 182 |
+
- `configs/grpo_small.yaml`
|
| 183 |
+
|
| 184 |
+
The training scaffold is intentionally minimal until the environment/verifier behavior is stable. Trackio metric names and GRPO defaults follow the project brief.
|
| 185 |
+
|
| 186 |
`training/train_grpo.py` in this repo is a config helper only; it does not execute training locally.
|
| 187 |
Use the Modal launchers in `scripts/modal_train_grpo.py` (persistent) and
|
| 188 |
`scripts/modal_ephemeral_train.py` (smoke) for real GRPO runs.
|
| 189 |
|
| 190 |
+
### Run SFT And GRPO Training Scripts
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 191 |
|
| 192 |
+
Training runs on Modal. Do not run the GRPO loop directly on the local machine;
|
| 193 |
+
use the launcher scripts so scenario cache preflight, Trackio logging, Modal
|
| 194 |
+
volumes, and Hub uploads stay consistent.
|
| 195 |
|
| 196 |
+
First install the Modal extra and prepare the scenario cache:
|
|
|
|
|
|
|
| 197 |
|
| 198 |
```bash
|
| 199 |
uv sync --extra modal
|
| 200 |
+
uv run --extra modal modal run scripts/modal_train_grpo.py --mode config
|
| 201 |
+
uv run --extra modal modal run scripts/modal_train_grpo.py --mode prepare-cache
|
| 202 |
```
|
| 203 |
|
| 204 |
+
Generate and verify SFT trajectories before supervised fine-tuning:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 205 |
|
| 206 |
```bash
|
| 207 |
uv run python scripts/generate_sft_dataset.py \
|
| 208 |
--teacher-model deepseek-ai/DeepSeek-V4-Pro \
|
| 209 |
--target-model unsloth/gemma-4-E2B-it \
|
| 210 |
--difficulty-levels 0,1,2,3 \
|
|
|
|
| 211 |
--episodes 75 \
|
| 212 |
--validation-episodes 20 \
|
| 213 |
--workers 8 \
|
| 214 |
--out-dir outputs/sft
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 215 |
|
|
|
|
|
|
|
|
|
|
| 216 |
uv run python scripts/generate_sft_dataset.py \
|
| 217 |
--verify-only \
|
| 218 |
--difficulty-levels 0,1,2,3 \
|
| 219 |
--out-dir outputs/sft
|
| 220 |
```
|
| 221 |
|
| 222 |
+
Run SFT on Modal and push the warm-start LoRA:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 223 |
|
| 224 |
```bash
|
| 225 |
uv run --extra modal modal run --detach scripts/modal_train_sft.py \
|
|
|
|
| 234 |
--detach
|
| 235 |
```
|
| 236 |
|
| 237 |
+
Continue with GRPO from the SFT adapter:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 238 |
|
| 239 |
```bash
|
| 240 |
uv run --extra modal modal run --detach scripts/modal_train_grpo.py \
|
| 241 |
--initial-adapter-repo-id Humanlearning/CyberSecurity_OWASP-unsloth-gemma-4-e2b-it-sft-lora \
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 242 |
--max-steps 300 \
|
| 243 |
--dataset-size 64 \
|
| 244 |
--num-generations 8 \
|
| 245 |
--max-completion-length 768 \
|
| 246 |
--difficulty 0 \
|
| 247 |
--trace-log-every 10 \
|
| 248 |
+
--trackio-space-id Humanlearning/CyberSecurity_OWASP-trackio \
|
| 249 |
+
--trackio-project CyberSecurity_OWASP-grpo \
|
| 250 |
--detach
|
| 251 |
```
|
| 252 |
|
| 253 |
+
For reward-rubric ablations, use the PowerShell launcher and configs under
|
| 254 |
+
`training/configs/reward_ablations/`:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 255 |
|
| 256 |
```powershell
|
| 257 |
+
.\scripts\launch_reward_ablations.ps1
|
| 258 |
```
|
| 259 |
|
| 260 |
+
Modal smoke and GRPO runs use `CYBERSECURITY_OWASP_SCENARIO_CACHE_MODE=require` and mount the persistent `CyberSecurity_OWASP-scenario-cache` volume. Prepare that cache before smoke/training:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 261 |
|
| 262 |
```bash
|
| 263 |
+
uv run --extra modal modal run scripts/modal_train_grpo.py --mode prepare-cache
|
| 264 |
+
uv run --extra modal modal run scripts/modal_ephemeral_train.py --mode prepare-cache
|
| 265 |
+
```
|
| 266 |
+
|
| 267 |
+
If the cache slice is missing or below the configured per-bucket minimum, Modal training fails before rollouts rather than compiling scenarios during the run.
|
| 268 |
+
The persistent GRPO launcher runs a CPU-only scenario-cache preflight before it starts the L4 GPU function, so missing cache coverage fails before GPU allocation.
|
| 269 |
+
|
| 270 |
+
## Trackio Run Tracking
|
| 271 |
+
|
| 272 |
+
Trackio is the default tracker for official runs. Set `TRACKIO_SPACE_ID` to log to a hosted Hugging Face Trackio Space; otherwise Trackio records locally.
|
| 273 |
+
|
| 274 |
+
```bash
|
| 275 |
+
export TRACKIO_SPACE_ID=<hf-user>/CyberSecurity_OWASP-trackio
|
| 276 |
+
export TRACKIO_PROJECT=CyberSecurity_OWASP-grpo
|
| 277 |
+
```
|
| 278 |
+
|
| 279 |
+
Use the tracked smoke wrapper instead of invoking pytest directly when producing run artifacts:
|
| 280 |
+
|
| 281 |
+
```bash
|
| 282 |
+
bash scripts/smoke_test.sh
|
| 283 |
+
uv run python scripts/track_pytest.py tests
|
| 284 |
+
```
|
| 285 |
+
|
| 286 |
+
Evaluation summaries saved through `training.eval_before_after.save_eval_summary(...)`, Modal smoke runs, and GRPO training configs all initialize Trackio runs with CyberSecurity_OWASP run names.
|
| 287 |
+
|
| 288 |
+
Training, baseline, and smoke runs also log the effective reward config at step
|
| 289 |
+
0. In Trackio, open **Media & Tables** and select the `reward_config` table to
|
| 290 |
+
see the actual values for each reward key, including stage-specific values,
|
| 291 |
+
caps, thresholds, terminate flags, and descriptions. Scalar metrics under
|
| 292 |
+
`reward_config/<key>/<field>` expose the same numeric values for plotting and
|
| 293 |
+
filtering, for example `reward_config/policy_inspected/value` and
|
| 294 |
+
`reward_config/shaping_weight/resolved`.
|
| 295 |
+
|
| 296 |
+
Each run config includes `reward_config_id`, `reward_config_hash`,
|
| 297 |
+
`reward_config_source`, `reward_mode`, and `reward_stage`. For manual ablations,
|
| 298 |
+
compare runs with the same scenario/model settings and different
|
| 299 |
+
`reward_config_hash` values to see which reward weights produced each training
|
| 300 |
+
curve.
|
| 301 |
+
|
| 302 |
+
## Modal Ephemeral Runs
|
| 303 |
+
|
| 304 |
+
Modal Labs support is kept in a separate launcher script so the local OpenEnv server and core training scaffold stay unchanged.
|
| 305 |
+
|
| 306 |
+
Install the optional local Modal client:
|
| 307 |
+
|
| 308 |
+
```bash
|
| 309 |
+
uv sync --extra modal
|
| 310 |
+
```
|
| 311 |
+
|
| 312 |
+
Run a temporary Modal app for a cheap environment/training smoke check:
|
| 313 |
+
|
| 314 |
+
```bash
|
| 315 |
+
uv run --extra modal modal run scripts/modal_ephemeral_train.py --mode prepare-cache
|
| 316 |
+
uv run --extra modal modal run scripts/modal_ephemeral_train.py --mode smoke --episodes 4
|
| 317 |
+
```
|
| 318 |
+
|
| 319 |
+
The app is ephemeral: Modal starts it for the command and stops it when the command exits. The remote result is written locally under `outputs/rollouts/` and the summary metrics are logged to Trackio.
|
| 320 |
+
|
| 321 |
+
You can also validate the GRPO config construction remotely:
|
| 322 |
+
|
| 323 |
+
```bash
|
| 324 |
+
uv run --extra modal modal run scripts/modal_ephemeral_train.py --mode grpo-config
|
| 325 |
+
```
|
| 326 |
+
|
| 327 |
+
The shell wrapper is equivalent:
|
| 328 |
+
|
| 329 |
+
```bash
|
| 330 |
+
MODE=smoke EPISODES=4 uv run --extra modal bash scripts/modal_run_ephemeral.sh
|
| 331 |
+
```
|
| 332 |
+
|
| 333 |
+
## Synthetic SFT Before GRPO
|
| 334 |
+
|
| 335 |
+
Use supervised fine-tuning to warm-start `unsloth/gemma-4-E2B-it` before GRPO.
|
| 336 |
+
The SFT generator executes every teacher action in the real environment and
|
| 337 |
+
keeps only trajectories that pass the deterministic reward verifier.
|
| 338 |
+
|
| 339 |
+
Generate a 300-train-episode curriculum SFT dataset across levels `0,1,2,3`:
|
| 340 |
+
|
| 341 |
+
```bash
|
| 342 |
+
uv run python scripts/generate_sft_dataset.py \
|
| 343 |
+
--teacher-model deepseek-ai/DeepSeek-V4-Pro \
|
| 344 |
+
--target-model unsloth/gemma-4-E2B-it \
|
| 345 |
+
--difficulty-levels 0,1,2,3 \
|
| 346 |
+
--difficulty-buckets 4 \
|
| 347 |
+
--episodes 75 \
|
| 348 |
+
--validation-episodes 20 \
|
| 349 |
+
--workers 8 \
|
| 350 |
+
--out-dir outputs/sft
|
| 351 |
+
```
|
| 352 |
+
|
| 353 |
+
`--episodes` is per difficulty level when `--difficulty-levels` is set, so
|
| 354 |
+
`--episodes 75` across four levels gives 300 total train episodes. Expect
|
| 355 |
+
roughly 2,400-4,500 chat-format JSONL rows because each successful trajectory
|
| 356 |
+
contributes one row per action step. The script writes JSONL rows under
|
| 357 |
+
`outputs/sft/`, trajectory artifacts under `outputs/sft/trajectories/`, a
|
| 358 |
+
dataset card at `outputs/sft/README.md`, and `outputs/sft/manifest.json` with
|
| 359 |
+
reward summaries and curriculum coverage.
|
| 360 |
+
|
| 361 |
+
Verify reward metadata before any training run:
|
| 362 |
+
|
| 363 |
+
```bash
|
| 364 |
+
uv run python scripts/generate_sft_dataset.py \
|
| 365 |
+
--verify-only \
|
| 366 |
+
--difficulty-levels 0,1,2,3 \
|
| 367 |
+
--out-dir outputs/sft
|
| 368 |
+
```
|
| 369 |
+
|
| 370 |
+
Push the verified dataset to Hugging Face Hub:
|
| 371 |
+
|
| 372 |
+
```bash
|
| 373 |
+
uv run python scripts/generate_sft_dataset.py \
|
| 374 |
+
--push-only \
|
| 375 |
+
--difficulty-levels 0,1,2,3 \
|
| 376 |
+
--out-dir outputs/sft \
|
| 377 |
+
--dataset-repo-id Humanlearning/CyberSecurity_OWASP-sft-dataset
|
| 378 |
+
```
|
| 379 |
+
|
| 380 |
+
The canonical dataset repo name is
|
| 381 |
+
`Humanlearning/CyberSecurity_OWASP-sft-dataset`. The upload is refused if
|
| 382 |
+
reward verification fails or `HF_TOKEN` is missing.
|
| 383 |
+
|
| 384 |
+
You can also generate and push in one command by adding `--push-to-hub` to the
|
| 385 |
+
generation command.
|
| 386 |
+
|
| 387 |
+
For local CI or smoke checks, add `--dry-run-oracle`; official SFT data should
|
| 388 |
+
use the teacher path and still pass the verifier gate above.
|
| 389 |
+
|
| 390 |
+
Launch SFT on Modal after reward verification passes:
|
| 391 |
+
|
| 392 |
+
```bash
|
| 393 |
+
uv run --extra modal modal run --detach scripts/modal_train_sft.py \
|
| 394 |
+
--local-train-path outputs/sft/train.jsonl \
|
| 395 |
+
--local-validation-path outputs/sft/validation.jsonl \
|
| 396 |
+
--local-manifest-path outputs/sft/manifest.json \
|
| 397 |
+
--required-difficulties 0,1,2,3 \
|
| 398 |
+
--trackio-space-id Humanlearning/CyberSecurity_OWASP-trackio \
|
| 399 |
+
--trackio-project CyberSecurity_OWASP-sft \
|
| 400 |
+
--output-repo-id Humanlearning/CyberSecurity_OWASP-unsloth-gemma-4-e2b-it-sft-lora \
|
| 401 |
+
--push-to-hub \
|
| 402 |
+
--detach
|
| 403 |
+
```
|
| 404 |
+
|
| 405 |
+
`scripts/modal_train_sft.py` re-checks the JSONL reward metadata locally before
|
| 406 |
+
upload and again inside Modal before loading the model. It refuses to start SFT
|
| 407 |
+
unless all required curriculum difficulties are represented and the verifier
|
| 408 |
+
reward metadata passes. The default SFT config trains the full dataset
|
| 409 |
+
(`--max-steps -1`) with bf16/tf32, LoRA rank 32, and Modal GPU fallback
|
| 410 |
+
`H200 -> H100 -> A100-80GB -> L40S`. TRL does not support packing or
|
| 411 |
+
assistant-only loss for the Gemma 4 vision-language loader, so both remain
|
| 412 |
+
disabled for this model. The script pre-tokenizes the small JSONL dataset
|
| 413 |
+
serially before constructing `SFTTrainer`, which avoids TRL multiprocessing
|
| 414 |
+
around the Gemma/Unsloth config object. It also uses the base Transformers loss
|
| 415 |
+
path to avoid a TRL entropy-metric incompatibility with Gemma 4 lazy logits. A
|
| 416 |
+
warm run for the 300-400 episode dataset should usually finish in about 20-60
|
| 417 |
+
minutes; first image or model-cache builds can push that closer to 45-90
|
| 418 |
+
minutes.
|
| 419 |
+
|
| 420 |
+
Continue GRPO from the SFT LoRA:
|
| 421 |
+
|
| 422 |
+
The GRPO launcher downloads the Hub adapter, attaches a matching trainable
|
| 423 |
+
Unsloth LoRA to Gemma 4, and then loads the adapter safetensors. This keeps the
|
| 424 |
+
SFT handoff compatible with Gemma 4's Unsloth linear wrappers.
|
| 425 |
+
|
| 426 |
+
```bash
|
| 427 |
+
uv run --extra modal modal run --detach scripts/modal_train_grpo.py \
|
| 428 |
+
--initial-adapter-repo-id Humanlearning/CyberSecurity_OWASP-unsloth-gemma-4-e2b-it-sft-lora \
|
| 429 |
+
--max-steps 300 \
|
| 430 |
+
--dataset-size 64 \
|
| 431 |
+
--num-generations 8 \
|
| 432 |
+
--difficulty 0 \
|
| 433 |
+
--trace-log-every 10 \
|
| 434 |
+
--detach
|
| 435 |
+
```
|
| 436 |
+
|
| 437 |
+
## Modal GRPO Training
|
| 438 |
+
|
| 439 |
+
The persistent GPU training launcher packages this local repo into Modal, trains
|
| 440 |
+
a small LoRA GRPO run, logs metrics and traces to Trackio, stores checkpoints in
|
| 441 |
+
the `CyberSecurity_OWASP-grpo-runs` Modal volume, and pushes the output adapter
|
| 442 |
+
to Hugging Face Hub.
|
| 443 |
+
|
| 444 |
+
Create a Modal secret named `CyberSecurity_OWASP-secrets` with `HF_TOKEN`, then
|
| 445 |
+
run the import/config check:
|
| 446 |
+
|
| 447 |
+
```bash
|
| 448 |
+
uv run --extra modal modal run scripts/modal_train_grpo.py --mode config
|
| 449 |
+
```
|
| 450 |
+
|
| 451 |
+
Run the default smoke GRPO job:
|
| 452 |
+
|
| 453 |
+
```bash
|
| 454 |
+
uv run --extra modal modal run scripts/modal_train_grpo.py --mode prepare-cache
|
| 455 |
+
uv run --extra modal modal run scripts/modal_train_grpo.py \
|
| 456 |
+
--max-steps 10 \
|
| 457 |
+
--dataset-size 16 \
|
| 458 |
+
--num-generations 6 \
|
| 459 |
+
--difficulty 0
|
| 460 |
+
```
|
| 461 |
+
|
| 462 |
+
For GPU-utilization tuning on the same single L4, start with a larger but still
|
| 463 |
+
bounded no-code trial:
|
| 464 |
+
|
| 465 |
+
```bash
|
| 466 |
+
uv run --extra modal modal run scripts/modal_train_grpo.py \
|
| 467 |
+
--max-steps 30 \
|
| 468 |
+
--dataset-size 64 \
|
| 469 |
+
--num-generations 8 \
|
| 470 |
+
--max-completion-length 256 \
|
| 471 |
+
--difficulty 0
|
| 472 |
+
```
|
| 473 |
+
|
| 474 |
+
The launcher exposes GRPO throughput knobs for follow-up trials:
|
| 475 |
+
|
| 476 |
+
```bash
|
| 477 |
+
# larger generation group, no vLLM
|
| 478 |
+
uv run --extra modal modal run scripts/modal_train_grpo.py \
|
| 479 |
+
--max-steps 30 --dataset-size 64 --num-generations 8 \
|
| 480 |
+
--max-completion-length 256 --trace-log-every 5
|
| 481 |
+
|
| 482 |
+
# vLLM colocate on the same L4
|
| 483 |
+
uv run --extra modal modal run scripts/modal_train_grpo.py \
|
| 484 |
+
--max-steps 30 --dataset-size 64 --num-generations 8 \
|
| 485 |
+
--max-completion-length 256 --use-vllm \
|
| 486 |
+
--vllm-gpu-memory-utilization 0.35 --trace-log-every 5
|
| 487 |
+
|
| 488 |
+
# larger microbatch if the vLLM trial does not OOM
|
| 489 |
+
uv run --extra modal modal run scripts/modal_train_grpo.py \
|
| 490 |
+
--max-steps 30 --dataset-size 64 --num-generations 8 \
|
| 491 |
+
--per-device-train-batch-size 2 --gradient-accumulation-steps 4 \
|
| 492 |
+
--max-completion-length 256 --use-vllm \
|
| 493 |
+
--vllm-gpu-memory-utilization 0.45 --trace-log-every 5
|
| 494 |
+
```
|
| 495 |
+
|
| 496 |
+
`per_device_train_batch_size * gradient_accumulation_steps * world_size` must
|
| 497 |
+
be divisible by `num_generations`; the launcher validates this before the GPU
|
| 498 |
+
container starts. Scalar Trackio metrics still log every reward callback, while
|
| 499 |
+
sample trace tables and Trace objects are throttled by `--trace-log-every`
|
| 500 |
+
(`1` restores every-callback logging, `0` disables trace artifacts).
|
| 501 |
+
|
| 502 |
+
### Parallel Modal GRPO Runs
|
| 503 |
+
|
| 504 |
+
Parallel Modal GRPO runs are safe when each run has its own seed range, run
|
| 505 |
+
name, and output target, while the shared cache volumes remain read-only.
|
| 506 |
+
Before launching another job, check what is already active:
|
| 507 |
+
|
| 508 |
+
```bash
|
| 509 |
+
uv run --extra modal modal app list
|
| 510 |
+
uv run --extra modal modal app logs <app-id>
|
| 511 |
+
```
|
| 512 |
+
|
| 513 |
+
Launch long-running parallel jobs with both Modal CLI detach and the launcher
|
| 514 |
+
detach flag. The CLI-level `--detach` keeps the remote function alive after the
|
| 515 |
+
local entrypoint exits; the launcher `--detach` prevents the parent Modal
|
| 516 |
+
function from waiting on the GPU call.
|
| 517 |
+
|
| 518 |
+
```bash
|
| 519 |
+
uv run --extra modal modal run --detach scripts/modal_train_grpo.py \
|
| 520 |
+
--max-steps 300 \
|
| 521 |
+
--dataset-size 64 \
|
| 522 |
+
--num-generations 8 \
|
| 523 |
+
--max-completion-length 768 \
|
| 524 |
+
--difficulty 0 \
|
| 525 |
+
--trace-log-every 10 \
|
| 526 |
+
--seed-start 10000 \
|
| 527 |
+
--detach
|
| 528 |
+
```
|
| 529 |
+
|
| 530 |
+
For multiple concurrent experiments:
|
| 531 |
+
|
| 532 |
+
- Use a unique `--seed-start` range for every run, normally spaced by at least
|
| 533 |
+
10,000 seeds.
|
| 534 |
+
- Keep `CYBERSECURITY_OWASP_SCENARIO_CACHE_MODE=require`; do not compile
|
| 535 |
+
scenarios during training.
|
| 536 |
+
- Do not run `prepare-cache --cache-force` while training jobs are active.
|
| 537 |
+
- Keep `--push-to-hub` disabled unless each run has a unique
|
| 538 |
+
`--output-repo-id`.
|
| 539 |
+
- Let the launcher generate unique timestamped Trackio run names, or set an
|
| 540 |
+
explicit `RUN_NAME` only when it is globally unique.
|
| 541 |
+
- Use the same Trackio Space/project for comparable metrics, but never reuse a
|
| 542 |
+
run name.
|
| 543 |
+
- Treat `CyberSecurity_OWASP-model-cache` and
|
| 544 |
+
`CyberSecurity_OWASP-scenario-cache` as shared read-mostly infrastructure
|
| 545 |
+
during training. Run outputs and checkpoints should stay under each run's
|
| 546 |
+
unique output directory.
|
| 547 |
+
|
| 548 |
+
If a Windows shell fails with a Unicode `charmap` encoding error during Modal
|
| 549 |
+
startup, rerun with UTF-8 enabled for that command:
|
| 550 |
+
|
| 551 |
+
```powershell
|
| 552 |
+
$env:PYTHONIOENCODING='utf-8'; $env:PYTHONUTF8='1'; uv run --extra modal modal run --detach scripts/modal_train_grpo.py --max-steps 300 --dataset-size 64 --num-generations 4 --max-completion-length 768 --difficulty 0 --trace-log-every 10 --seed-start 60000 --detach
|
| 553 |
+
```
|
| 554 |
+
|
| 555 |
+
If running from a public repository and you do not want Modal to package the
|
| 556 |
+
local workspace, use public source mode:
|
| 557 |
+
|
| 558 |
+
```bash
|
| 559 |
+
uv run --extra modal modal run scripts/modal_train_grpo.py \
|
| 560 |
+
--source-mode public \
|
| 561 |
+
--repo-url https://github.com/humandotlearning/CyberSecurity_OWASP.git \
|
| 562 |
+
--repo-branch master \
|
| 563 |
+
--max-steps 10 \
|
| 564 |
+
--dataset-size 16 \
|
| 565 |
+
--num-generations 6 \
|
| 566 |
+
--difficulty 0
|
| 567 |
+
```
|
| 568 |
+
|
| 569 |
+
Defaults are derived from `HF_TOKEN`:
|
| 570 |
+
|
| 571 |
+
- Trackio Space: `<hf-user>/CyberSecurity_OWASP-trackio`
|
| 572 |
+
- Trackio project: `CyberSecurity_OWASP-grpo`
|
| 573 |
+
- Training model: `unsloth/gemma-4-E2B-it`
|
| 574 |
+
- Output repo: `<hf-user>/CyberSecurity_OWASP-unsloth-gemma-4-e2b-it-grpo-lora`
|
| 575 |
+
|
| 576 |
+
Override these with `--trackio-space-id`, `--trackio-project`, and
|
| 577 |
+
`--output-repo-id` when needed. The persistent GRPO launcher intentionally rejects non-Gemma model overrides so smoke runs match the Unsloth Gemma 4 E2B RL notebook.
|
| 578 |
+
|
| 579 |
+
## Docker / Spaces
|
| 580 |
+
|
| 581 |
+
```bash
|
| 582 |
+
docker build -t CyberSecurity_OWASP:latest -f server/Dockerfile .
|
| 583 |
+
docker run --rm -p 8000:8000 CyberSecurity_OWASP:latest
|
| 584 |
+
openenv push --repo-id <username>/CyberSecurity_OWASP
|
| 585 |
+
```
|