feat: update evaluation results and training scripts for Qwen2.5-7B-Instruct

Updated evaluation metrics and added new JSONL files for the Qwen2.5-7B-Instruct model, including detailed task performance data. Enhanced training scripts to reflect changes in output paths and parameters, improving the training process for GRPO and SFT. Updated visualizations and metadata to ensure comprehensive tracking of model performance across various tasks.

results/runs/Qwen2.5-7B-Instruct__base-qwen2.5-7b/base-qwen2.5-7b_eval.jsonl

@@ -0,0 +1,3 @@
+{"task_id": "e1_launch_readiness", "policy_kind": "model", "steps": 3, "total_reward": 1.39, "terminal_reward": 0.91, "reward_trace": [0.29000000000000004, 0.19, 0.9099999999999999], "verifier_pass_rate": 1.0, "passed_checks": ["qa_report_present", "final_rec_valid", "no_missed_milestones"], "failed_checks": [], "milestones_total": 2, "milestones_complete": 2, "milestones_missed": 0, "invalid_action_count": 0, "env_error_count": 0, "errors": [], "final_swd_version": 4, "success": true, "episode_index": 0, "model_stage": "base-qwen2.5-7b", "policy": "hf", "model": "Qwen/Qwen2.5-7B-Instruct", "adapter": ""}
+{"task_id": "m1_budget_reallocation", "policy_kind": "model", "steps": 6, "total_reward": 2.056667, "terminal_reward": 0.806667, "reward_trace": [0.09000000000000001, 0.29000000000000004, 0.19, 0.39, 0.29000000000000004, 0.8066666666666666], "verifier_pass_rate": 0.666667, "passed_checks": ["required_agents_consulted", "conflict_logged", "conflict_resolved", "phased_plan"], "failed_checks": ["budget_constraint_acknowledged", "reasoning_documented"], "milestones_total": 3, "milestones_complete": 3, "milestones_missed": 0, "invalid_action_count": 0, "env_error_count": 0, "errors": [], "final_swd_version": 7, "success": false, "episode_index": 0, "model_stage": "base-qwen2.5-7b", "policy": "hf", "model": "Qwen/Qwen2.5-7B-Instruct", "adapter": ""}
+{"task_id": "h1_acquisition_defence", "policy_kind": "model", "steps": 8, "total_reward": 2.283333, "terminal_reward": 0.753333, "reward_trace": [0.09000000000000001, 0.09000000000000001, 0.29000000000000004, 0.19, 0.39, 0.29000000000000004, 0.19, 0.7533333333333332], "verifier_pass_rate": 0.666667, "passed_checks": ["all_agents_consulted", "conflict_explicitly_resolved", "resolution_has_type", "counter_offer_present", "deadline_present", "retention_addressed", "no_single_agent_copied", "swd_version_rich"], "failed_checks": ["multi_conflict_logged", "rich_reasoning_log", "timeline_constraint_acknowledged", "all_phases_reached"], "milestones_total": 5, "milestones_complete": 3, "milestones_missed": 0, "invalid_action_count": 0, "env_error_count": 0, "errors": [], "final_swd_version": 9, "success": false, "episode_index": 0, "model_stage": "base-qwen2.5-7b", "policy": "hf", "model": "Qwen/Qwen2.5-7B-Instruct", "adapter": ""}

results/runs/Qwen2.5-7B-Instruct__base-qwen2.5-7b/metadata.json

@@ -0,0 +1,32 @@
+{
+  "model_stage": "base-qwen2.5-7b",
+  "policy": "hf",
+  "model": "Qwen/Qwen2.5-7B-Instruct",
+  "adapter": "",
+  "tasks": [
+    "e1_launch_readiness",
+    "m1_budget_reallocation",
+    "h1_acquisition_defence"
+  ],
+  "episodes": 1,
+  "max_steps": 30,
+  "max_new_tokens": 1536,
+  "eval_file": "results/runs/Qwen2.5-7B-Instruct__base-qwen2.5-7b/base-qwen2.5-7b_eval.jsonl",
+  "metrics_by_task": {
+    "e1_launch_readiness": {
+      "avg_terminal_reward": 0.91,
+      "avg_verifier_pass_rate": 1.0,
+      "success_rate": 1.0
+    },
+    "m1_budget_reallocation": {
+      "avg_terminal_reward": 0.806667,
+      "avg_verifier_pass_rate": 0.666667,
+      "success_rate": 0.0
+    },
+    "h1_acquisition_defence": {
+      "avg_terminal_reward": 0.753333,
+      "avg_verifier_pass_rate": 0.666667,
+      "success_rate": 0.0
+    }
+  }
+}

results/runs/Qwen2.5-7B-Instruct__grpo_Qwen2.5-7B-Instruct__grpo-qwen2.5-7b/grpo-qwen2.5-7b_eval.jsonl

@@ -0,0 +1,3 @@
+{"task_id": "e1_launch_readiness", "policy_kind": "model", "steps": 4, "total_reward": 1.58, "terminal_reward": 0.91, "reward_trace": [0.29000000000000004, 0.19, 0.19, 0.9099999999999999], "verifier_pass_rate": 1.0, "passed_checks": ["qa_report_present", "final_rec_valid", "no_missed_milestones"], "failed_checks": [], "milestones_total": 2, "milestones_complete": 2, "milestones_missed": 0, "invalid_action_count": 0, "env_error_count": 0, "errors": [], "final_swd_version": 5, "success": true, "episode_index": 0, "model_stage": "grpo-qwen2.5-7b", "policy": "hf", "model": "Qwen/Qwen2.5-7B-Instruct", "adapter": "outputs/grpo_Qwen2.5-7B-Instruct"}
+{"task_id": "m1_budget_reallocation", "policy_kind": "model", "steps": 5, "total_reward": 1.508333, "terminal_reward": 0.548333, "reward_trace": [0.09000000000000001, 0.29000000000000004, 0.19, 0.39, 0.5483333333333333], "verifier_pass_rate": 0.5, "passed_checks": ["required_agents_consulted", "conflict_logged", "conflict_resolved"], "failed_checks": ["phased_plan", "budget_constraint_acknowledged", "reasoning_documented"], "milestones_total": 3, "milestones_complete": 2, "milestones_missed": 0, "invalid_action_count": 0, "env_error_count": 0, "errors": [], "final_swd_version": 6, "success": false, "episode_index": 0, "model_stage": "grpo-qwen2.5-7b", "policy": "hf", "model": "Qwen/Qwen2.5-7B-Instruct", "adapter": "outputs/grpo_Qwen2.5-7B-Instruct"}
+{"task_id": "h1_acquisition_defence", "policy_kind": "model", "steps": 60, "total_reward": 1.54, "terminal_reward": -0.01, "reward_trace": [0.09000000000000001, 0.09000000000000001, 0.29000000000000004, 0.19, 0.19, 0.39, 0.29000000000000004, -0.03, -0.03, -0.03, 0.19, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, 0.19, 0.19, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01, -0.01], "verifier_pass_rate": 0.416667, "passed_checks": ["all_agents_consulted", "conflict_explicitly_resolved", "resolution_has_type", "no_single_agent_copied", "swd_version_rich"], "failed_checks": ["multi_conflict_logged", "rich_reasoning_log", "counter_offer_present", "deadline_present", "retention_addressed", "timeline_constraint_acknowledged", "all_phases_reached"], "milestones_total": 5, "milestones_complete": 2, "milestones_missed": 3, "invalid_action_count": 47, "env_error_count": 47, "errors": ["invalid_action: 1 validation error for CorpAction\nreward\n  Extra inputs are not permitted [type=extra_forbidden, input_value=0.39, input_type=float]\n    For further information visit https://errors.pydantic.dev/2.12/v/extra_forbidden", "invalid_action: 1 validation error for CorpAction\nreward\n  Extra inputs are not permitted [type=extra_forbidden, input_value=0.39, input_type=float]\n    For further information visit https://errors.pydantic.dev/2.12/v/extra_forbidden", "invalid_action: 1 validation error for CorpAction\nreward\n  Extra inputs are not permitted [type=extra_forbidden, input_value=0.39, input_type=float]\n    For further information visit https://errors.pydantic.dev/2.12/v/extra_forbidden", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)", "invalid_action: Expecting ',' delimiter: line 1 column 78 (char 77)"], "final_swd_version": 14, "success": false, "episode_index": 0, "model_stage": "grpo-qwen2.5-7b", "policy": "hf", "model": "Qwen/Qwen2.5-7B-Instruct", "adapter": "outputs/grpo_Qwen2.5-7B-Instruct"}

results/runs/Qwen2.5-7B-Instruct__grpo_Qwen2.5-7B-Instruct__grpo-qwen2.5-7b/metadata.json

@@ -0,0 +1,32 @@
+{
+  "model_stage": "grpo-qwen2.5-7b",
+  "policy": "hf",
+  "model": "Qwen/Qwen2.5-7B-Instruct",
+  "adapter": "outputs/grpo_Qwen2.5-7B-Instruct",
+  "tasks": [
+    "e1_launch_readiness",
+    "m1_budget_reallocation",
+    "h1_acquisition_defence"
+  ],
+  "episodes": 1,
+  "max_steps": 30,
+  "max_new_tokens": 1536,
+  "eval_file": "results/runs/Qwen2.5-7B-Instruct__grpo_Qwen2.5-7B-Instruct__grpo-qwen2.5-7b/grpo-qwen2.5-7b_eval.jsonl",
+  "metrics_by_task": {
+    "e1_launch_readiness": {
+      "avg_terminal_reward": 0.91,
+      "avg_verifier_pass_rate": 1.0,
+      "success_rate": 1.0
+    },
+    "m1_budget_reallocation": {
+      "avg_terminal_reward": 0.548333,
+      "avg_verifier_pass_rate": 0.5,
+      "success_rate": 0.0
+    },
+    "h1_acquisition_defence": {
+      "avg_terminal_reward": -0.01,
+      "avg_verifier_pass_rate": 0.416667,
+      "success_rate": 0.0
+    }
+  }
+}

results/runs/Qwen2.5-7B-Instruct__sft_Qwen2.5-7B-Instruct__sft-qwen2.5-7b/metadata.json

@@ -0,0 +1,32 @@
+{
+  "model_stage": "sft-qwen2.5-7b",
+  "policy": "hf",
+  "model": "Qwen/Qwen2.5-7B-Instruct",
+  "adapter": "outputs/sft_Qwen2.5-7B-Instruct",
+  "tasks": [
+    "e1_launch_readiness",
+    "m1_budget_reallocation",
+    "h1_acquisition_defence"
+  ],
+  "episodes": 1,
+  "max_steps": 30,
+  "max_new_tokens": 1536,
+  "eval_file": "results/runs/Qwen2.5-7B-Instruct__sft_Qwen2.5-7B-Instruct__sft-qwen2.5-7b/sft-qwen2.5-7b_eval.jsonl",
+  "metrics_by_task": {
+    "e1_launch_readiness": {
+      "avg_terminal_reward": 0.91,
+      "avg_verifier_pass_rate": 1.0,
+      "success_rate": 1.0
+    },
+    "m1_budget_reallocation": {
+      "avg_terminal_reward": 0.943333,
+      "avg_verifier_pass_rate": 1.0,
+      "success_rate": 1.0
+    },
+    "h1_acquisition_defence": {
+      "avg_terminal_reward": 0.515,
+      "avg_verifier_pass_rate": 0.5,
+      "success_rate": 0.0
+    }
+  }
+}

results/runs/Qwen2.5-7B-Instruct__sft_Qwen2.5-7B-Instruct__sft-qwen2.5-7b/sft-qwen2.5-7b_eval.jsonl

@@ -0,0 +1,3 @@
+{"task_id": "e1_launch_readiness", "policy_kind": "model", "steps": 4, "total_reward": 1.58, "terminal_reward": 0.91, "reward_trace": [0.29000000000000004, 0.19, 0.19, 0.9099999999999999], "verifier_pass_rate": 1.0, "passed_checks": ["qa_report_present", "final_rec_valid", "no_missed_milestones"], "failed_checks": [], "milestones_total": 2, "milestones_complete": 2, "milestones_missed": 0, "invalid_action_count": 0, "env_error_count": 0, "errors": [], "final_swd_version": 5, "success": true, "episode_index": 0, "model_stage": "sft-qwen2.5-7b", "policy": "hf", "model": "Qwen/Qwen2.5-7B-Instruct", "adapter": "outputs/sft_Qwen2.5-7B-Instruct"}
+{"task_id": "m1_budget_reallocation", "policy_kind": "model", "steps": 6, "total_reward": 2.093333, "terminal_reward": 0.943333, "reward_trace": [0.09000000000000001, 0.29000000000000004, 0.19, 0.19, 0.39, 0.9433333333333334], "verifier_pass_rate": 1.0, "passed_checks": ["required_agents_consulted", "conflict_logged", "conflict_resolved", "phased_plan", "budget_constraint_acknowledged", "reasoning_documented"], "failed_checks": [], "milestones_total": 3, "milestones_complete": 3, "milestones_missed": 0, "invalid_action_count": 0, "env_error_count": 0, "errors": [], "final_swd_version": 7, "success": true, "episode_index": 0, "model_stage": "sft-qwen2.5-7b", "policy": "hf", "model": "Qwen/Qwen2.5-7B-Instruct", "adapter": "outputs/sft_Qwen2.5-7B-Instruct"}
+{"task_id": "h1_acquisition_defence", "policy_kind": "model", "steps": 15, "total_reward": 2.915, "terminal_reward": 0.515, "reward_trace": [0.09000000000000001, 0.09000000000000001, 0.29000000000000004, 0.19, 0.19, 0.39, 0.29000000000000004, -0.03, -0.03, -0.03, 0.39, 0.19, 0.19, 0.19, 0.515], "verifier_pass_rate": 0.5, "passed_checks": ["all_agents_consulted", "multi_conflict_logged", "conflict_explicitly_resolved", "resolution_has_type", "no_single_agent_copied", "swd_version_rich"], "failed_checks": ["rich_reasoning_log", "counter_offer_present", "deadline_present", "retention_addressed", "timeline_constraint_acknowledged", "all_phases_reached"], "milestones_total": 5, "milestones_complete": 3, "milestones_missed": 0, "invalid_action_count": 0, "env_error_count": 0, "errors": [], "final_swd_version": 16, "success": false, "episode_index": 0, "model_stage": "sft-qwen2.5-7b", "policy": "hf", "model": "Qwen/Qwen2.5-7B-Instruct", "adapter": "outputs/sft_Qwen2.5-7B-Instruct"}

training/train_grpo.py

@@ -10,14 +10,26 @@ Run on Colab, Lightning AI H100, or another GPU machine after SFT:
 
   python training/train_grpo.py \
     --model Qwen/Qwen2.5-7B-Instruct \
-    --adapter outputs/sft_adapter \
+    --adapter outputs/sft_Qwen2.5-7B-Instruct \
     --examples data/processed/e1_m1_clean.jsonl \
-    --output outputs/grpo_adapter \
-    --max-steps 30
+    --output outputs/grpo_Qwen2.5-7B-Instruct \
+    --max-steps 50 \
+    --push-to-hub your-org/corp-gym-grpo-qwen2.5-7b
 
 The reward function recreates the environment state from a verified action prefix,
 applies the sampled next action, and returns the real environment reward plus
 penalties for invalid JSON/actions.
+
+Optional speedups (Unsloth picks these up automatically when importable):
+
+- **Flash Attention 2** (`flash_attn`): largest win for long contexts / GRPO rollouts.
+  Install a wheel matching your **exact** `torch` and CUDA build, or compile with
+  `CUDA_HOME` pointing at the **same** CUDA version PyTorch was built for (check
+  `python -c "import torch; print(torch.version.cuda)"`). Mismatch (e.g. nvcc 13.0
+  vs torch cu128) breaks the build; fix the toolkit or use an image with FA2 preinstalled.
+- **xFormers**: already used as a fallback when FA2 is missing; still slower than FA2.
+- **Qwen3.x “linear attention” fast path**: only for those architectures; not used for
+  Qwen2.5 GRPO. See Unsloth logs if you train Qwen3.5+.
 """
 
 from __future__ import annotations
@@ -144,18 +156,39 @@ def environment_reward(
 def main() -> None:
     parser = argparse.ArgumentParser(description="Train CORP-ENV GRPO adapter.")
     parser.add_argument("--model", default="Qwen/Qwen2.5-7B-Instruct")
-    parser.add_argument("--adapter", default="outputs/sft_adapter")
+    parser.add_argument("--adapter", default="outputs/sft_Qwen2.5-7B-Instruct")
     parser.add_argument("--examples", default="data/processed/e1_m1_clean.jsonl")
-    parser.add_argument("--output", default="outputs/grpo_adapter")
+    parser.add_argument("--output", default="outputs/grpo_Qwen2.5-7B-Instruct")
     parser.add_argument("--tasks", default="e1_launch_readiness,m1_budget_reallocation")
-    parser.add_argument("--repeats", type=int, default=128)
+    parser.add_argument(
+        "--repeats",
+        type=int,
+        default=32,
+        help="When --examples is missing/empty, build this many synthetic oracle prefixes per task. Ignored if the JSONL yields rows.",
+    )
     parser.add_argument("--max-prompt-length", type=int, default=8192)
     parser.add_argument("--max-completion-length", type=int, default=1024)
     parser.add_argument("--lr", type=float, default=5e-6)
     parser.add_argument("--batch-size", type=int, default=1)
     parser.add_argument("--grad-accum", type=int, default=8)
-    parser.add_argument("--generations", type=int, default=4)
-    parser.add_argument("--max-steps", type=int, default=150)
+    parser.add_argument(
+        "--generations",
+        type=int,
+        default=2,
+        help="GRPO samples per prompt per step. Lower = faster steps, noisier gradients (try 2 for quick runs, 4 for fuller RL).",
+    )
+    parser.add_argument(
+        "--max-steps",
+        type=int,
+        default=50,
+        help="Optimizer steps (not env episodes). Lower finishes sooner; raise for stronger fit.",
+    )
+    parser.add_argument(
+        "--save-steps",
+        type=int,
+        default=None,
+        help="Checkpoint every N steps. Default: min(25, max(5, max_steps//2)) so short runs still save.",
+    )
     parser.add_argument("--optim", default="adamw_8bit")
     parser.add_argument("--push-to-hub", default="")
     args = parser.parse_args()
@@ -164,10 +197,10 @@ def main() -> None:
     os.environ.setdefault("CORP_DISABLE_LLM_JUDGE", "1")
 
     try:
+        from unsloth import FastLanguageModel, PatchFastRL
         from datasets import Dataset
         from peft import PeftModel
         from trl import GRPOConfig, GRPOTrainer
-        from unsloth import FastLanguageModel, PatchFastRL
     except ImportError as exc:
         raise SystemExit(
             "GRPO training requires unsloth, trl, datasets, and peft. On Lightning AI, install with:\n"
@@ -186,6 +219,15 @@ def main() -> None:
         dtype=None,
         load_in_4bit=True,
     )
+    if getattr(tokenizer, "pad_token", None) is None and getattr(
+        tokenizer, "eos_token", None
+    ) is not None:
+        tokenizer.pad_token = tokenizer.eos_token
+
+    save_steps = args.save_steps
+    if save_steps is None:
+        save_steps = min(25, max(5, args.max_steps // 2))
+
     if args.adapter:
         model = PeftModel.from_pretrained(model, args.adapter, is_trainable=True)
     else:
@@ -210,7 +252,7 @@ def main() -> None:
         max_completion_length=args.max_completion_length,
         max_steps=args.max_steps,
         logging_steps=5,
-        save_steps=25,
+        save_steps=save_steps,
         save_total_limit=3,
         optim=args.optim,
         bf16=True,