feat: add IFBench training configuration

configs/ifbench/README.md

@@ -0,0 +1,61 @@
+# IFBench / IF-RLVR Training Runbook
+
+This directory adds a minimal `prime-rl` training configuration for the IFBench / IF-RLVR bounty.
+
+## Source material
+
+- Bounty: IF-RLVR/Bench on Algora / Prime Intellect
+- Ai2 code: https://github.com/allenai/IFBench
+- IFBench data collection: https://huggingface.co/collections/allenai/ifbench-683f590687f61b512558cdf1
+- IF-RLVR training data named by Ai2: `allenai/IF_multi_constraints_upto5`
+- Prime Intellect Lab environment discovered publicly at: https://app.primeintellect.ai/dashboard/environments/shashwat23/ifbench
+
+## Configuration
+
+The training config is:
+
+```text
+configs/ifbench/rl.toml
+```
+
+It follows the same lightweight pattern as `configs/gsm8k/rl.toml` and `configs/hendrycks_math/rl.toml`, but uses the Prime Lab registry environment ID:
+
+```toml
+[[orchestrator.train.env]]
+id = "shashwat23/ifbench"
+name = "ifbench"
+```
+
+In `prime-rl`, slash-qualified environment IDs are installed before training via:
+
+```bash
+uv run --no-sync prime env install shashwat23/ifbench
+```
+
+## Run
+
+From the repo root:
+
+```bash
+uv run prime-rl rl configs/ifbench/rl.toml
+```
+
+If the environment is not already installed locally, the orchestrator should install `shashwat23/ifbench` automatically before loading it.
+
+## Validation done here
+
+This contribution is intentionally small: it wires the IFBench environment into `prime-rl` as a train config/runbook. No heavy GPU training was run locally.
+
+Lightweight checks:
+
+```bash
+python3 - <<'PY'
+import tomllib, pathlib
+tomllib.loads(pathlib.Path('configs/ifbench/rl.toml').read_text())
+print('ok configs/ifbench/rl.toml')
+PY
+```
+
+## Maintainer confirmation needed
+
+Please confirm whether `shashwat23/ifbench` is the intended canonical environment ID for the bounty. If the final registry name differs, only the `id` in `configs/ifbench/rl.toml` should need changing.

configs/ifbench/rl.toml

@@ -0,0 +1,27 @@
+max_steps = 100
+seq_len = 2048
+
+[wandb]
+project = "ifbench-debug"
+name = "ifbench"
+
+[model]
+name = "PrimeIntellect/Qwen3-0.6B"
+
+[orchestrator]
+batch_size = 512
+rollouts_per_example = 16
+
+[orchestrator.train.sampling]
+max_completion_tokens = 2048
+
+[[orchestrator.train.env]]
+# Prime Intellect Lab environment discovered at:
+# https://app.primeintellect.ai/dashboard/environments/shashwat23/ifbench
+# The slash-qualified ID triggers `prime env install` in prime-rl before training.
+id = "shashwat23/ifbench"
+name = "ifbench"
+
+[trainer] # Default trainer config
+
+[inference] # Default inference config