Add configs/kaggle_grpo.yaml

configs/kaggle_grpo.yaml

@@ -1,4 +1,4 @@
-# Kaggle GRPO defaults — T4 x2, 4-bit LoRA, eval-before-checkpoint
+# Kaggle agent GRPO — T4 x2, 4-bit LoRA, deepagents SWE harness + remote Docker grader
 accelerator: t4x2
 seed: 42
 
@@ -9,31 +9,33 @@ model:
     - Qwen/Qwen2.5-1.5B-Instruct
     - HuggingFaceTB/SmolLM2-1.7B-Instruct
     - google/gemma-2-2b-it
-  max_seq_length: 2048
+  max_seq_length: 4096
   load_in_4bit: true
   lora_r: 16
   lora_alpha: 32
   lora_dropout: 0.0
 
 data:
-  hf_dataset_id: vanishingradient/deepagents-rl-mbpp
-  train_split: train
-  eval_split: eval
+  hf_dataset_id: vanishingradient/deepagents-rl-swe-lite
+  train_split: dev
+  eval_split: test
   prompt_column: prompt
+  max_train_instances: 16
 
 grpo:
   learning_rate: 1.0e-6
-  num_generations: 4
-  max_completion_length: 512
+  num_generations: 2
+  max_completion_length: 1024
+  max_tool_calling_iterations: 12
   per_device_train_batch_size: 1
   gradient_accumulation_steps: 4
-  max_steps: 40
+  max_steps: 20
   logging_steps: 1
-  # Save every N optimizer steps; ALWAYS run held-out eval + Langfuse BEFORE save
   save_steps: 5
   eval_before_checkpoint:
     enabled: true
-    n_problems: 16
+    n_problems: 2
+    max_eval_turns: 8
     temperature: 0.0
     max_new_tokens: 512
   push_to_hub: true
@@ -41,5 +43,11 @@ grpo:
 
 observability:
   wandb_project: deepagents-rl-coding
-  langfuse_ compulsory: true
+  wandb_run_name: kaggle-grpo-swe
+  langfuse_compulsory: true
   langfuse_eval_trace_name: kaggle-ckpt-eval
+
+grader:
+  # Set via secrets: SWE_GRADER_URL (cloudflared/ngrok -> local swe_grader_server)
+  # optional SWE_GRADER_TOKEN
+  required: true