{
  "generated_at": "2026-08-11T00:00:00Z",
  "selection_version": 1,
  "selection_dimensions": [
    "measured_result_strength",
    "agent_training_relevance",
    "reproducibility",
    "pedagogical_surprise",
    "claim_integrity"
  ],
  "experiments": [
    {
      "rank": 1,
      "slug": "echo-lfm-three-seeds",
      "title": "OpenEnv ECHO with LFM2.5-350M",
      "kind": "positive_result",
      "publishable": true,
      "model": "LiquidAI/LFM2.5-350M",
      "method": "ECHO environment-token prediction",
      "primary_metric": {"name": "heldout_env_ce", "before": 13.741455078125, "after": 0.41339555382728577},
      "models": [
        "https://huggingface.co/burtenshaw/openenv-echo-world-model-liquidai-lfm2.5-350m-seed0",
        "https://huggingface.co/burtenshaw/openenv-echo-world-model-liquidai-lfm2.5-350m-seed1",
        "https://huggingface.co/burtenshaw/openenv-echo-world-model-liquidai-lfm2.5-350m-seed2"
      ],
      "trackio": "https://huggingface.co/spaces/burtenshaw/openenv-examples-trackio",
      "keywords": ["echo", "world model", "openenv", "environment tokens", "lfm", "verifier free"]
    },
    {
      "rank": 2,
      "slug": "qwen-mbpp-sft",
      "title": "Qwen3-0.6B MBPP SFT",
      "kind": "positive_result",
      "publishable": true,
      "model": "Qwen/Qwen3-0.6B",
      "method": "TRL completion-only SFT LoRA",
      "primary_metric": {"name": "mbpp_accuracy", "before": 0.2607003891050584, "after": 0.3852140077821012},
      "models": ["https://huggingface.co/burtenshaw/qwen3-0.6b-mbpp-sft-lr1e4-r16-best-len2k"],
      "trackio": "https://huggingface.co/spaces/burtenshaw/mbpp-sft-trackio",
      "keywords": ["mbpp", "qwen", "coding", "sft", "lora", "completion only"]
    },
    {
      "rank": 3,
      "slug": "lfm-commonsenseqa-sft",
      "title": "LFM2.5-350M CommonsenseQA SFT",
      "kind": "positive_result",
      "publishable": true,
      "model": "LiquidAI/LFM2.5-350M",
      "method": "TRL completion-only SFT LoRA",
      "primary_metric": {"name": "forced_choice_accuracy", "before": 0.20393120393120392, "after": 0.25225225225225223},
      "models": ["https://huggingface.co/burtenshaw/lfm25-350m-commonsenseqa-sft-lr1e4-r16-2k"],
      "trackio": "https://huggingface.co/spaces/burtenshaw/lfm25-commonsenseqa-sft-trackio",
      "keywords": ["commonsenseqa", "commonsense", "lfm", "sft", "forced choice"]
    },
    {
      "rank": 4,
      "slug": "browsergym-click-test",
      "title": "BrowserGym tool-call training",
      "kind": "scoped_smoke_result",
      "publishable": true,
      "model": "LiquidAI/LFM2.5-230M",
      "method": "bounded LoRA tool-call training",
      "primary_metric": {"name": "click_test_reward", "before": 0.0, "after": 1.0},
      "models": ["https://huggingface.co/burtenshaw/openenv-browsergym-lfm25-230m-reward-improved"],
      "trackio": "https://huggingface.co/spaces/burtenshaw/trackio-browsergym-lfm25-230m-20260629-105734",
      "keywords": ["browsergym", "browser", "miniwob", "tool call", "click test", "openenv"]
    },
    {
      "rank": 5,
      "slug": "toolcalc-sft-vs-grpo",
      "title": "Toolcalc SFT versus GRPO",
      "kind": "controlled_negative",
      "publishable": true,
      "method": "SFT and GRPO reward comparison",
      "primary_metric": {"name": "exact_answer_success", "sft": 0.5, "grpo": 0.0},
      "models": [],
      "trackio": "https://huggingface.co/spaces/burtenshaw/training-agents-toolcalc-trackio",
      "keywords": ["toolcalc", "grpo", "sft", "reward hacking", "tool calling", "arithmetic"]
    },
    {
      "rank": 6,
      "slug": "tinker-echo-fallback",
      "title": "Tinker ECHO preflight and local fallback",
      "kind": "scoped_positive_result",
      "publishable": true,
      "model": "distilgpt2",
      "method": "Tinker datum test plus local ECHO fallback",
      "primary_metric": {"name": "heldout_env_ce", "before": 6.182, "after": 0.271},
      "models": ["https://huggingface.co/burtenshaw/openenv-echo-world-model-distilgpt2-seed0"],
      "trackio": "https://huggingface.co/spaces/burtenshaw/agent-handover-trackio",
      "keywords": ["tinker", "echo", "distilgpt2", "fallback", "world model"]
    },
    {
      "rank": 7,
      "slug": "onpolicy-gkd-soft-vs-hard",
      "title": "Qwen/Gemma on-policy GKD",
      "kind": "controlled_negative",
      "publishable": true,
      "model": "Qwen/Qwen3.5-4B",
      "method": "TRL on-policy GKD",
      "primary_metric": {"name": "strict_proxy_success", "before": 0.25, "after": 0.25},
      "secondary_metric": {"name": "valid_command_rate", "before": 0.3684210526, "after": 0.6842105263},
      "models": [],
      "trackio": "https://huggingface.co/spaces/burtenshaw/agent-handover-trackio",
      "keywords": ["gkd", "distillation", "on policy", "qwen", "gemma", "valid command"]
    },
    {
      "rank": 8,
      "slug": "qlora-loss-vs-behavior",
      "title": "QLoRA SFT loss versus behavior",
      "kind": "controlled_negative",
      "publishable": true,
      "model": "Qwen/Qwen2.5-0.5B-Instruct",
      "method": "12-step 4-bit QLoRA SFT",
      "primary_metric": {"name": "eval_loss", "before": 2.51088, "after": 2.191612},
      "secondary_metric": {"name": "mean_token_accuracy", "before": 0.49868, "after": 0.535078},
      "models": [],
      "trackio": "https://huggingface.co/spaces/burtenshaw/agent-handover-trackio",
      "keywords": ["qlora", "sft", "instruction following", "behavior", "loss", "token accuracy"]
    },
    {
      "rank": 9,
      "slug": "gemma-pi-mono-transfer",
      "title": "Gemma Pi-Mono transfer",
      "kind": "controlled_negative",
      "publishable": true,
      "model": "google/gemma-4-E2B-it",
      "method": "TRL SFT LoRA on agent traces",
      "primary_metric": {"name": "humaneval", "before": 0.7256, "after": 0.744},
      "secondary_metric": {"name": "mbpp", "before": 0.716, "after": 0.651},
      "models": [],
      "trackio": "https://huggingface.co/spaces/burtenshaw/pi-mono-sft-trackio",
      "keywords": ["pi mono", "gemma", "humaneval", "mbpp", "transfer", "specialization"]
    },
    {
      "rank": 10,
      "slug": "awm-session-stress",
      "title": "Agent World Model session stress",
      "kind": "environment_reliability",
      "publishable": true,
      "method": "bounded concurrent OpenEnv session stress",
      "primary_metric": {"name": "session_success_rate", "before": null, "after": 1.0},
      "models": [],
      "trackio": "https://huggingface.co/spaces/burtenshaw/agent-handover-trackio",
      "keywords": ["agent world model", "awm", "openenv", "latency", "concurrency", "sessions"]
    }
  ]
}