{ "$schema": "./schema/models.v1.schema.json", "schemaVersion": "1.0.0", "manifestId": "slime-models-2026-07-12", "observationSize": 12, "actions": [ "NOOP", "FORWARD", "FORWARD_JUMP", "JUMP", "BACKWARD_JUMP", "BACKWARD" ], "models": [ { "id": "plain", "name": "Plain DQN", "shortName": "Plain DQN", "family": "ablation", "trainingMethod": "DQN without Dueling, Double or N-step", "role": "candidate", "runtime": "onnx", "outputKind": "q_values", "officialEligible": true, "featured": false, "hero": false, "url": "model_plain.onnx?v=79fecd5bb93a", "sha256": "79fecd5bb93a3dbc1f4d4043d44d23221a6cb01b7ae4ce88ba7a6e2cd1b06ba5", "sizeBytes": 108545, "checkpoint": "dqn_plain_vec.pt", "timesteps": 5000000, "description": "Plain DQN ablation used to isolate the value of later algorithmic components." }, { "id": "dueling", "name": "Dueling DQN", "shortName": "Dueling", "family": "ablation", "trainingMethod": "Dueling DQN", "role": "candidate", "runtime": "onnx", "outputKind": "q_values", "officialEligible": true, "featured": false, "hero": false, "url": "model.onnx?v=ad50f1da3565", "sha256": "ad50f1da356581af7bea35fe537069d0dfeb918780b09801c3d8260dc1470cae", "sizeBytes": 143007, "checkpoint": "dqn_vec.pt", "timesteps": 5000000, "description": "Dueling value/advantage architecture without Double DQN or N-step returns." }, { "id": "duel_double", "name": "Dueling + Double DQN", "shortName": "Duel + Double", "family": "ablation", "trainingMethod": "Dueling Double DQN", "role": "candidate", "runtime": "onnx", "outputKind": "q_values", "officialEligible": true, "featured": false, "hero": false, "url": "model_duel_double.onnx?v=ed61e06b2280", "sha256": "ed61e06b22808acf0faff20076f91663a70d17e22b6dde82357b6505f1ac77a5", "sizeBytes": 143007, "checkpoint": "dqn_v1_dd_vec.pt", "timesteps": 5000000, "description": "Ablation combining the Dueling head with Double DQN target selection." }, { "id": "duel_nstep", "name": "Dueling + N-step", "shortName": "Duel + N-step", "family": "ablation", "trainingMethod": "Dueling DQN with 3-step returns", "role": "candidate", "runtime": "onnx", "outputKind": "q_values", "officialEligible": true, "featured": false, "hero": false, "url": "model_duel_nstep.onnx?v=57d9b7e24e82", "sha256": "57d9b7e24e828056da8f6b1cc370aefc5b77737eec51a65f04c08730e70d0e98", "sizeBytes": 143007, "checkpoint": "dqn_v1_dn_vec.pt", "timesteps": 5000000, "description": "Ablation testing N-step returns without Double DQN correction." }, { "id": "v1", "name": "V1 Cold", "shortName": "V1 Cold", "family": "v1", "trainingMethod": "Dueling Double DQN with N-step returns", "role": "candidate", "runtime": "onnx", "outputKind": "q_values", "officialEligible": true, "featured": false, "hero": false, "url": "model_v1.onnx?v=d48c283dbc77", "sha256": "d48c283dbc77677347d3f0683e3b064662f7e18785b4b0097b74aeab4b8306ef", "sizeBytes": 143007, "checkpoint": "dqn_v1_vec.pt", "timesteps": 5000000, "description": "Cold-start V1 policy trained without a self-play opponent pool." }, { "id": "v1_selfplay", "name": "V1 Self-play (legacy)", "shortName": "V1 Self-play", "family": "v1", "trainingMethod": "V1 snapshot self-play, legacy baseline contract", "role": "candidate", "runtime": "onnx", "outputKind": "q_values", "officialEligible": true, "featured": false, "hero": false, "url": "model_v1_selfplay.onnx?v=86d901bae293", "sha256": "86d901bae2932601771e0a8aeb39722755d2e5c4bc1446f35d106226201234a3", "sizeBytes": 143007, "checkpoint": "dqn_v1_vec_selfplay.pt", "timesteps": 20000000, "description": "Legacy self-play lineage retained to show the ecosystem effect found during the study." }, { "id": "v1_purepool", "name": "V1 Pure-pool", "shortName": "V1 Pure-pool", "family": "v1", "trainingMethod": "V1 self-play without baseline mixing", "role": "candidate", "runtime": "onnx", "outputKind": "q_values", "officialEligible": true, "featured": false, "hero": false, "url": "model_v1_purepool.onnx?v=9aae62a06370", "sha256": "9aae62a06370287d95d959211bbbe5bb6aa12ef7a67f0f33afda74524ab27871", "sizeBytes": 143007, "checkpoint": "dqn_v1_vec_selfplay_purepool.pt", "timesteps": 20000000, "description": "Self-play ablation with no legacy reference policy mixed into the opponent pool." }, { "id": "v1_selfplay_seeded", "name": "V1 Self-play Seeded", "shortName": "V1 Seeded", "family": "v1", "trainingMethod": "V1 snapshot self-play with seeded opponents", "role": "candidate", "runtime": "onnx", "outputKind": "q_values", "officialEligible": true, "featured": false, "hero": false, "url": "model_v1_selfplay_seeded.onnx?v=74fba88c841a", "sha256": "74fba88c841afab89f88c5df832f5cbc338d91424258a56d52e17d21bf46ea07", "sizeBytes": 143007, "checkpoint": "dqn_v1_vec_selfplay_seeded.pt", "timesteps": 20000000, "description": "Legacy V1 self-play policy initialized with a broader fixed opponent set." }, { "id": "v1_sp_fixed", "name": "Trained V1", "shortName": "V1 Hero", "family": "v1", "trainingMethod": "V1 snapshot self-play with corrected baseline mixing", "role": "candidate", "runtime": "onnx", "outputKind": "q_values", "officialEligible": true, "featured": true, "hero": true, "url": "model_v1_selfplay_fixed.onnx?v=fd18bedfaf27", "sha256": "fd18bedfaf27a6cf5065332236eeb9d65be18a3ae4b4fbe194186793f0170d33", "sizeBytes": 143007, "checkpoint": "dqn_v1_vec_selfplay_fixed.pt", "timesteps": 20000000, "description": "The more active exhibition hero, trained with corrected self-play orientation and baseline inputs." }, { "id": "ppo", "name": "PPO (legacy)", "shortName": "PPO", "family": "ppo", "trainingMethod": "PPO self-play", "role": "candidate", "runtime": "onnx", "outputKind": "logits", "officialEligible": true, "featured": false, "hero": false, "url": "model_ppo.onnx?v=386926592f45", "sha256": "386926592f45cac888d2ee27e2eb45becd96a0a6382f8a74ced5133a75fdebd1", "sizeBytes": 76704, "checkpoint": "ppo_selfplay.pt", "timesteps": 5000000, "description": "Early PPO policy retained as a historical comparison." }, { "id": "ppo_fixed", "name": "PPO Low-entropy", "shortName": "PPO Low-ent", "family": "ppo", "trainingMethod": "PPO self-play with corrected baseline inputs", "role": "candidate", "runtime": "onnx", "outputKind": "logits", "officialEligible": true, "featured": false, "hero": false, "url": "model_ppo_fixed.onnx?v=a311893d50df", "sha256": "a311893d50df6bda7668533ed686b043b620d6163846677388b30f7e021a91df", "sizeBytes": 76704, "checkpoint": "ppo_lowent_20m_fixed.pt", "timesteps": 20000000, "description": "A low-entropy PPO run that illustrates premature policy collapse." }, { "id": "ppo_rescue", "name": "PPO Rescue", "shortName": "PPO Rescue", "family": "ppo", "trainingMethod": "PPO self-play with entropy and stability rescue", "role": "candidate", "runtime": "onnx", "outputKind": "logits", "officialEligible": true, "featured": false, "hero": false, "url": "model_ppo_rescue.onnx?v=9227f5cda983", "sha256": "9227f5cda98374f4802b13f0f8ebdc29d6dbac956298bd06ca4b6ec50636bcef", "sizeBytes": 76704, "checkpoint": "ppo_rescue_20m.pt", "timesteps": 20000000, "description": "Higher-entropy PPO policy recovered from the collapse observed in the low-entropy run." }, { "id": "rainbow", "name": "Rainbow-lite 20M (legacy)", "shortName": "Rainbow 20M", "family": "rainbow", "trainingMethod": "Rainbow-lite self-play", "role": "candidate", "runtime": "onnx", "outputKind": "q_values", "officialEligible": true, "featured": false, "hero": false, "url": "model_rainbow.onnx?v=f09a7385a7d0", "sha256": "f09a7385a7d0f589d907dbbac662f7adcfa41a563e964db762c4609c7c1330f7", "sizeBytes": 143007, "checkpoint": "dqn_rainbow_lite_selfplay.pt", "timesteps": 20000000, "description": "Legacy Rainbow-lite self-play policy kept for historical comparison." }, { "id": "rainbow_5m", "name": "Rainbow-lite 5M", "shortName": "Rainbow 5M", "family": "rainbow", "trainingMethod": "Rainbow-lite without self-play", "role": "candidate", "runtime": "onnx", "outputKind": "q_values", "officialEligible": true, "featured": false, "hero": false, "url": "model_rainbow_5m.onnx?v=195d96859fda", "sha256": "195d96859fdacde3cd80046afa858b56b86417f97c2671f7b01e038f63dd4923", "sizeBytes": 143007, "checkpoint": "dqn_rainbow_lite_5m.pt", "timesteps": 5000000, "description": "Five-million-step Rainbow-lite policy trained without self-play." }, { "id": "rainbow_5m_sp", "name": "Rainbow-lite 5M Self-play", "shortName": "Rainbow 5M SP", "family": "rainbow", "trainingMethod": "Rainbow-lite snapshot self-play", "role": "candidate", "runtime": "onnx", "outputKind": "q_values", "officialEligible": true, "featured": false, "hero": false, "url": "model_rainbow_5m_sp.onnx?v=1ad1363b7b7f", "sha256": "1ad1363b7b7f8f6c141005c8357f1f2fbbbea353a3ccfa1285a7c76c7c0011fc", "sizeBytes": 143007, "checkpoint": "dqn_rainbow_lite_5m_selfplay.pt", "timesteps": 5000000, "description": "Five-million-step Rainbow-lite policy with snapshot self-play." }, { "id": "rainbow_sp_fixed", "name": "Trained Rainbow", "shortName": "Rainbow Hero", "family": "rainbow", "trainingMethod": "Aggression-shaped Rainbow-lite self-play", "role": "candidate", "runtime": "onnx", "outputKind": "q_values", "officialEligible": true, "featured": true, "hero": true, "url": "model_rainbow_sp_fixed.onnx?v=252e43ca5a4b", "sha256": "252e43ca5a4b0e5b600f1cef9d1e164868fddce912ba4bdb3e8305c213fad13c", "sizeBytes": 143007, "checkpoint": "rainbow_aggro2_seed0.pt", "timesteps": 4600000, "description": "Release policy selected on a fixed both-sides benchmark after aggression-shaped fine-tuning." }, { "id": "rainbow_full", "name": "Rainbow Full", "shortName": "Rainbow Full", "family": "rainbow", "trainingMethod": "Dueling Double N-step PER C51 NoisyNet", "role": "candidate", "runtime": "onnx", "outputKind": "q_values", "officialEligible": true, "featured": false, "hero": false, "url": "model_rainbow_full.onnx?v=ed5bc87bd9f7", "sha256": "ed5bc87bd9f7aba8f0a106598041ed064e75bb1a7c4f159b44c1bcfd4487ecf2", "sizeBytes": 234965, "checkpoint": "dqn_rainbow_full.pt", "timesteps": 5000000, "description": "Full Rainbow ablation; C51 support mismatch made this variant an instructive failure." }, { "id": "rainbow_eps", "name": "Rainbow C51", "shortName": "Rainbow C51", "family": "rainbow", "trainingMethod": "Dueling Double N-step PER C51", "role": "candidate", "runtime": "onnx", "outputKind": "q_values", "officialEligible": true, "featured": false, "hero": false, "url": "model_rainbow_eps.onnx?v=951635898172", "sha256": "9516358981727699ea06f3a48b628a1be0dec57deb7c8d066a7877a11ec15fe0", "sizeBytes": 234917, "checkpoint": "dqn_rainbow_eps.pt", "timesteps": 5000000, "description": "Distributional Rainbow ablation without NoisyNet." }, { "id": "baseline", "name": "Legacy Neuroevolution RNN", "shortName": "RNN Reference", "family": "reference", "trainingMethod": "2015 self-play neuroevolution", "role": "reference", "runtime": "js-rnn", "outputKind": "actions", "officialEligible": false, "featured": true, "hero": false, "url": null, "sha256": null, "sizeBytes": null, "checkpoint": "slimevolleygym BaselinePolicy", "timesteps": null, "description": "A learned 112-scalar recurrent policy from the original Neural Slime Volleyball project, used only as a fixed reference." } ] }