Run Status

• Status: complete_skipped
• Adapter present: True
• Latest checkpoint: outputs/qwen-capability-light/stage2-capability-step-sft/checkpoint-260
• Best checkpoint: outputs/qwen-capability-light/stage2-capability-step-sft/checkpoint-260
• Best eval loss: 2.5278091430664062
• Trainer state: outputs/qwen-capability-light/stage2-capability-step-sft/trainer_state.json
• Global step: 260
• First Loss: 1.5216938257217407
• Final Loss: 1.655211091041565
• Min Loss: 0.6422638297080994
• Max Loss: 2.0936222076416016
• Loss Points: 260
• First Eval Loss: 2.6287848949432373
• Final Eval Loss: 2.5278091430664062
• Min Eval Loss: 2.5278091430664062
• Max Eval Loss: 2.6287848949432373
• Eval Loss Points: 14
• Best Eval Loss: 2.5278091430664062
• Best Global Step: 260
• Train Runtime S: 3794.8465

Generated files:

• training_config.json
• stage_report.json
• loss_history.csv
• loss_curve.svg
• eval_loss_history.csv
• eval_loss_curve.svg

Context

• Purpose: Capability next-action SFT on coding-agent decisions plus verifier-style gap repairs.
• Previous adapter: armand0e/qwen3.5-capability-light-behavior-seed-lora
• Next stage: stage3-capability-dpo
• Base model: Qwen/Qwen3.5-2B
• Data file: data/assembled/sft_qwen_next_actions_capability_light.jsonl
• Eval file: data/eval/eval_next_actions_with_retention.jsonl
• LoRA r/alpha/dropout: 16 / 16 / 0.0
• Learning rate: 8e-07
• Epochs: 1.0
• Merged 16-bit model: armand0e/qwen3.5-capability-light-capability-step-merged-16bit

Upstream Data

• armand0e/qwen3.7-max-pi-traces
• armand0e/badlogicgames-pi-mono-opus-filtered
• armand0e/gpt-5.5-agent
• armand0e/gpt-5.5-chat
• TeichAI/claude-4.5-opus-high-reasoning-250x
• TeichAI/Claude-Opus-4.6-Reasoning-887x

Compact Local Sample

json
Copy code
{
  "messages": [
    {
      "content": "User/task context:\nuser: Give only the answer: (98 + 459) - 34 = ?",
      "role": "user"
    },
    {
      "content": "523",
      "reasoning_content": "Expression evaluates to 523. The answer should return only the numeric answer, while making sure no extra words.",
      "role": "assistant"
    }
  ],
  "metadata": {
    "expected": 523,
    "family": "arithmetic",
    "key": "gap/arithmetic/00000",
    "source": "gap_capability_pack"
  },
  "source": "gap_capability_pack"
}

Reproduction

The exact stage command and package versions are in training_config.json.