webshop_single_expert_qwen2_5_3_step_300_seed0: LoRA adapters (steps 100/200/300, best_val) + manifest/README; first 3B run, below 1.5B reference

training/webshop_single_expert_qwen2_5_3_step_300_seed0/README.md

@@ -0,0 +1,15 @@
+# webshop_single_expert_qwen2_5_3_step_300_seed0
+
+WebShop single-expert LoRA training of Qwen2.5-**3B**-Instruct, seed 0, terminal step 300, recipe aligned
+to `webshop_single_expert_qwen2_5_1_5_step_300_seed0` (only micro-batch sizes halved).
+
+Published checkpoints: steps 100, 200, 300 and `best_val` (selected at step 220, sampled
+validation success rate 0.7188).
+
+Official fixed-goal evaluation (goals 0..499 exactly once, greedy), run on the training host:
+step 300 -> success rate **0.650**, mean task score **0.764**. The 1.5B reference evaluated with the same
+script on the same host gives 0.682 / 0.821 (collaborator README: 0.668 / 0.784), so this first 3B run is
+**below** the 1.5B reference and is kept for reference only; the 3B recipe will be re-tuned.
+
+Periodic training validation used 128 sampled goals (temp 0.4) and is not the official benchmark; the
+per-step values and file hashes are in `manifest.json`.

training/webshop_single_expert_qwen2_5_3_step_300_seed0/checkpoints/best_val/actor/lora_adapter/adapter_config.json

@@ -0,0 +1,45 @@
+{
+    "task_type": "CAUSAL_LM",
+    "peft_type": "LORA",
+    "auto_mapping": null,
+    "base_model_name_or_path": "Qwen/Qwen2.5-3B-Instruct",
+    "revision": null,
+    "inference_mode": false,
+    "r": 64,
+    "target_modules": [
+        "down_proj",
+        "gate_proj",
+        "k_proj",
+        "o_proj",
+        "q_proj",
+        "up_proj",
+        "v_proj"
+    ],
+    "exclude_modules": null,
+    "lora_alpha": 64,
+    "lora_dropout": 0.0,
+    "fan_in_fan_out": false,
+    "bias": "none",
+    "use_rslora": false,
+    "modules_to_save": null,
+    "init_lora_weights": true,
+    "layers_to_transform": null,
+    "layers_pattern": null,
+    "rank_pattern": {},
+    "alpha_pattern": {},
+    "megatron_config": null,
+    "megatron_core": "megatron.core",
+    "trainable_token_indices": null,
+    "loftq_config": {},
+    "eva_config": null,
+    "corda_config": null,
+    "use_dora": false,
+    "use_qalora": false,
+    "qalora_group_size": 16,
+    "layer_replication": null,
+    "runtime_config": {
+        "ephemeral_gpu_offload": false
+    },
+    "lora_bias": false,
+    "target_parameters": null
+}

training/webshop_single_expert_qwen2_5_3_step_300_seed0/checkpoints/best_val/actor/lora_adapter/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c70bf7d74bc2f2f7239702aefc0a9478044af6339b9b3b72b32346c1a80cf568
+size 479005032

training/webshop_single_expert_qwen2_5_3_step_300_seed0/checkpoints/best_val_metric.json

@@ -0,0 +1 @@
+{"step": 220, "metric_name": "val/success_rate", "metric_value": 0.71875}

training/webshop_single_expert_qwen2_5_3_step_300_seed0/checkpoints/global_step_100/actor/lora_adapter/adapter_config.json

@@ -0,0 +1,45 @@
+{
+    "task_type": "CAUSAL_LM",
+    "peft_type": "LORA",
+    "auto_mapping": null,
+    "base_model_name_or_path": "Qwen/Qwen2.5-3B-Instruct",
+    "revision": null,
+    "inference_mode": false,
+    "r": 64,
+    "target_modules": [
+        "down_proj",
+        "gate_proj",
+        "k_proj",
+        "o_proj",
+        "q_proj",
+        "up_proj",
+        "v_proj"
+    ],
+    "exclude_modules": null,
+    "lora_alpha": 64,
+    "lora_dropout": 0.0,
+    "fan_in_fan_out": false,
+    "bias": "none",
+    "use_rslora": false,
+    "modules_to_save": null,
+    "init_lora_weights": true,
+    "layers_to_transform": null,
+    "layers_pattern": null,
+    "rank_pattern": {},
+    "alpha_pattern": {},
+    "megatron_config": null,
+    "megatron_core": "megatron.core",
+    "trainable_token_indices": null,
+    "loftq_config": {},
+    "eva_config": null,
+    "corda_config": null,
+    "use_dora": false,
+    "use_qalora": false,
+    "qalora_group_size": 16,
+    "layer_replication": null,
+    "runtime_config": {
+        "ephemeral_gpu_offload": false
+    },
+    "lora_bias": false,
+    "target_parameters": null
+}

training/webshop_single_expert_qwen2_5_3_step_300_seed0/checkpoints/global_step_100/actor/lora_adapter/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d8c876edd66842685aa7585c7cf5749bf11841fc06aefec28a897db2bf44d7c5
+size 479005032

training/webshop_single_expert_qwen2_5_3_step_300_seed0/checkpoints/global_step_200/actor/lora_adapter/adapter_config.json

@@ -0,0 +1,45 @@
+{
+    "task_type": "CAUSAL_LM",
+    "peft_type": "LORA",
+    "auto_mapping": null,
+    "base_model_name_or_path": "Qwen/Qwen2.5-3B-Instruct",
+    "revision": null,
+    "inference_mode": false,
+    "r": 64,
+    "target_modules": [
+        "down_proj",
+        "gate_proj",
+        "k_proj",
+        "o_proj",
+        "q_proj",
+        "up_proj",
+        "v_proj"
+    ],
+    "exclude_modules": null,
+    "lora_alpha": 64,
+    "lora_dropout": 0.0,
+    "fan_in_fan_out": false,
+    "bias": "none",
+    "use_rslora": false,
+    "modules_to_save": null,
+    "init_lora_weights": true,
+    "layers_to_transform": null,
+    "layers_pattern": null,
+    "rank_pattern": {},
+    "alpha_pattern": {},
+    "megatron_config": null,
+    "megatron_core": "megatron.core",
+    "trainable_token_indices": null,
+    "loftq_config": {},
+    "eva_config": null,
+    "corda_config": null,
+    "use_dora": false,
+    "use_qalora": false,
+    "qalora_group_size": 16,
+    "layer_replication": null,
+    "runtime_config": {
+        "ephemeral_gpu_offload": false
+    },
+    "lora_bias": false,
+    "target_parameters": null
+}

training/webshop_single_expert_qwen2_5_3_step_300_seed0/checkpoints/global_step_200/actor/lora_adapter/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8dde20918dbb4713f1e6de3ab1c4b8f377f9064b4a489954123ca810196edbb2
+size 479005032

training/webshop_single_expert_qwen2_5_3_step_300_seed0/checkpoints/global_step_300/actor/lora_adapter/adapter_config.json

@@ -0,0 +1,45 @@
+{
+    "task_type": "CAUSAL_LM",
+    "peft_type": "LORA",
+    "auto_mapping": null,
+    "base_model_name_or_path": "Qwen/Qwen2.5-3B-Instruct",
+    "revision": null,
+    "inference_mode": false,
+    "r": 64,
+    "target_modules": [
+        "down_proj",
+        "gate_proj",
+        "k_proj",
+        "o_proj",
+        "q_proj",
+        "up_proj",
+        "v_proj"
+    ],
+    "exclude_modules": null,
+    "lora_alpha": 64,
+    "lora_dropout": 0.0,
+    "fan_in_fan_out": false,
+    "bias": "none",
+    "use_rslora": false,
+    "modules_to_save": null,
+    "init_lora_weights": true,
+    "layers_to_transform": null,
+    "layers_pattern": null,
+    "rank_pattern": {},
+    "alpha_pattern": {},
+    "megatron_config": null,
+    "megatron_core": "megatron.core",
+    "trainable_token_indices": null,
+    "loftq_config": {},
+    "eva_config": null,
+    "corda_config": null,
+    "use_dora": false,
+    "use_qalora": false,
+    "qalora_group_size": 16,
+    "layer_replication": null,
+    "runtime_config": {
+        "ephemeral_gpu_offload": false
+    },
+    "lora_bias": false,
+    "target_parameters": null
+}

training/webshop_single_expert_qwen2_5_3_step_300_seed0/checkpoints/global_step_300/actor/lora_adapter/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f6c4f34a6410ea796200f8cae8b7d542f60cfb1cda7e1c989bafbace42ee09d7
+size 479005032

training/webshop_single_expert_qwen2_5_3_step_300_seed0/checkpoints/latest_checkpointed_iteration.txt

@@ -0,0 +1 @@
+300

training/webshop_single_expert_qwen2_5_3_step_300_seed0/manifest.json

@@ -0,0 +1,164 @@
+{
+  "experiment": "webshop_single_expert_qwen2_5_3_step_300_seed0",
+  "environment": "webshop",
+  "base_model": "Qwen/Qwen2.5-3B-Instruct",
+  "seed": 0,
+  "terminal_step": 300,
+  "recipe": "aligned to webshop_single_expert_qwen2_5_1_5_step_300_seed0 (GiGPO, nothinking, lr 3e-6, LoRA r64/a64, 16 goals x n=8, 300 steps, val every 20 steps on 128 sampled goals); only micro-batch sizes halved for memory",
+  "status": "first 3B attempt; below the 1.5B reference on the official eval -- kept for reference, to be re-tuned",
+  "best_val": {
+    "metric": "val/success_rate",
+    "step": 220,
+    "value": 0.71875
+  },
+  "periodic_validation_128goals_t0.4": [
+    {
+      "step": 0,
+      "val/success_rate": 0.0
+    },
+    {
+      "step": 20,
+      "val/success_rate": 0.18
+    },
+    {
+      "step": 40,
+      "val/success_rate": 0.328
+    },
+    {
+      "step": 60,
+      "val/success_rate": 0.508
+    },
+    {
+      "step": 80,
+      "val/success_rate": 0.547
+    },
+    {
+      "step": 100,
+      "val/success_rate": 0.594
+    },
+    {
+      "step": 120,
+      "val/success_rate": 0.609
+    },
+    {
+      "step": 140,
+      "val/success_rate": 0.664
+    },
+    {
+      "step": 160,
+      "val/success_rate": 0.68
+    },
+    {
+      "step": 180,
+      "val/success_rate": 0.688
+    },
+    {
+      "step": 200,
+      "val/success_rate": 0.711
+    },
+    {
+      "step": 220,
+      "val/success_rate": 0.719
+    },
+    {
+      "step": 240,
+      "val/success_rate": 0.68
+    },
+    {
+      "step": 260,
+      "val/success_rate": 0.719
+    },
+    {
+      "step": 280,
+      "val/success_rate": 0.703
+    },
+    {
+      "step": 300,
+      "val/success_rate": 0.711
+    }
+  ],
+  "official_eval_500goals_greedy": {
+    "global_step_300": {
+      "success_rate": 0.65,
+      "mean_task_score": 0.7641,
+      "host": "this server, scripts/eval/run_eval_single_dense_webshop_local.sh"
+    },
+    "reference_1_5B_step_300_same_host": {
+      "success_rate": 0.682,
+      "mean_task_score": 0.8207
+    },
+    "reference_1_5B_step_300_collaborator_readme": {
+      "success_rate": 0.668,
+      "mean_task_score": 0.784373
+    },
+    "best_val": "pending (being evaluated)"
+  },
+  "artifact_scope": {
+    "included": [
+      "LoRA adapters",
+      "checkpoint selection metadata"
+    ],
+    "excluded": [
+      "FSDP shards",
+      "optimizer state",
+      "rollout data",
+      "sampler snapshots",
+      "logs",
+      "TensorBoard",
+      "pruning results"
+    ]
+  },
+  "created": "2026-08-27",
+  "files": [
+    {
+      "path": "checkpoints/best_val/actor/lora_adapter/adapter_config.json",
+      "bytes": 1100,
+      "sha256": "46d31221c0fe3c5513068d707db17de57923a275bd7917db2664e9b6ab0eac97"
+    },
+    {
+      "path": "checkpoints/best_val/actor/lora_adapter/adapter_model.safetensors",
+      "bytes": 479005032,
+      "sha256": "c70bf7d74bc2f2f7239702aefc0a9478044af6339b9b3b72b32346c1a80cf568"
+    },
+    {
+      "path": "checkpoints/best_val_metric.json",
+      "bytes": 73,
+      "sha256": "d34c236ed3ef00a2d6b6221875b2f3b6ca9733cb060ce048a828d63f64f631f0"
+    },
+    {
+      "path": "checkpoints/global_step_100/actor/lora_adapter/adapter_config.json",
+      "bytes": 1100,
+      "sha256": "46d31221c0fe3c5513068d707db17de57923a275bd7917db2664e9b6ab0eac97"
+    },
+    {
+      "path": "checkpoints/global_step_100/actor/lora_adapter/adapter_model.safetensors",
+      "bytes": 479005032,
+      "sha256": "d8c876edd66842685aa7585c7cf5749bf11841fc06aefec28a897db2bf44d7c5"
+    },
+    {
+      "path": "checkpoints/global_step_200/actor/lora_adapter/adapter_config.json",
+      "bytes": 1100,
+      "sha256": "46d31221c0fe3c5513068d707db17de57923a275bd7917db2664e9b6ab0eac97"
+    },
+    {
+      "path": "checkpoints/global_step_200/actor/lora_adapter/adapter_model.safetensors",
+      "bytes": 479005032,
+      "sha256": "8dde20918dbb4713f1e6de3ab1c4b8f377f9064b4a489954123ca810196edbb2"
+    },
+    {
+      "path": "checkpoints/global_step_300/actor/lora_adapter/adapter_config.json",
+      "bytes": 1100,
+      "sha256": "46d31221c0fe3c5513068d707db17de57923a275bd7917db2664e9b6ab0eac97"
+    },
+    {
+      "path": "checkpoints/global_step_300/actor/lora_adapter/adapter_model.safetensors",
+      "bytes": 479005032,
+      "sha256": "f6c4f34a6410ea796200f8cae8b7d542f60cfb1cda7e1c989bafbace42ee09d7"
+    },
+    {
+      "path": "checkpoints/latest_checkpointed_iteration.txt",
+      "bytes": 3,
+      "sha256": "983bd614bb5afece5ab3b6023f71147cd7b6bc2314f9d27af7422541c6558389"
+    }
+  ]
+}