kissin42 commited on
Commit
544af0e
Β·
verified Β·
1 Parent(s): d31eb67

Publish DungeonEscape step 70,800 bundle

Browse files

Replace the DungeonEscape safetensors bundle with training step 70,800, trained using simple + medium + expert data. KV64 closed-loop validation over seeds 100..108: 187/324 agent successes, 101/108 cooperative-group successes, and 324/324 natural terminations. Legacy ONNX artifacts are retained and explicitly labeled in the model card.

README.md CHANGED
@@ -76,11 +76,10 @@ of history, plus the ONNX variants used for Unity deployment.
76
  | `crawler` | `Tuple(Box(126), Box(32))` | `Box(20, [-1, 1])` |
77
  | `pushblock` | `Tuple(Box(105), Box(105))` | `Discrete(7)` |
78
 
79
- The result column is the **ONNX at context length 32**, the recommended Unity
80
- deployment artifact: a self-contained, fixed-batch rolling-window graph derived
81
- from the sibling safetensors bundle. Running the bundle itself in PyTorch at the
82
- same length scores slightly higher β€” 63.89% and 41.25%. Both are broken out by
83
- length below.
84
 
85
  **Unity behavior names:** `DungeonEscape`, `SoccerTwos`. HF bundle subfolders are
86
  lowercase and case-sensitive.
@@ -90,8 +89,9 @@ Each bundle wraps its ego spaces in a named container β€”
90
  scene. `Batch` is the fixed ONNX batch: 36 for DungeonEscape (12 arenas Γ— 3
91
  agents), 16 for SoccerTwos (8 fields Γ— 2 controlled agents).
92
 
93
- Training data is expert-free: bundles are trained using the `simple-v0` and
94
- `medium-v0` tiers only; expert trajectories are not used for training.
 
95
 
96
  Metrics differ by scene type and are not comparable across rows. DungeonEscape is
97
  cooperative: **agent success** is the fraction of ego agents that finish with a
@@ -124,8 +124,8 @@ Results at 0.5Γ—, 1Γ—, and 2Γ— the training length:
124
 
125
  | Bundle | Runtime | 16 (0.5Γ—) | 32 (1Γ—) | 64 (2Γ—) |
126
  |---|---|---:|---:|---:|
127
- | `dungeon-escape` | PyTorch (cached) | 57.41% | **63.89%** | 61.11% |
128
- | `dungeon-escape` | ONNX (windowed) | 56.48% | **60.19%** | not published ‑ |
129
  | `soccer-twos` | PyTorch (cached) | β€” † | **41.25%** | β€” † |
130
  | `soccer-twos` | ONNX (windowed) | 40.00% | **40.00%** | 32.50% |
131
  | `pushblock` | PyTorch (cached) | 159/160 | 159/160 | 159/160 |
@@ -146,22 +146,32 @@ axis is not unmeasured for this bundle β€” only the cached runtime at 0.5Γ— and
146
  <details>
147
  <summary><strong>Per-environment protocol and full results</strong></summary>
148
 
149
- **DungeonEscape** β€” matching release-23 model-removed build, Unity environment
150
- seeds `100..102`, 36 ego agents per launch (12 cooperative groups of three),
151
- CUDA, time scale 40. Trained from the `simple-v0` and `medium-v0` tiers; the
152
- expert tier was not used.
153
 
154
- | Length | Runtime | Agent success | Cooperative-group success |
155
- |---:|---|---:|---:|
156
- | 16 | PyTorch (cached) | 62/108 (57.41%) | 34/36 (94.44%) |
157
- | 16 | ONNX (windowed) | 61/108 (56.48%) | 33/36 (91.67%) |
158
- | **32** | **PyTorch (cached)** | **69/108 (63.89%)** | **34/36 (94.44%)** |
159
- | **32** | **ONNX (windowed)** | **65/108 (60.19%)** | **34/36 (94.44%)** |
160
- | 64 | PyTorch (cached) | 66/108 (61.11%) | 33/36 (91.67%) |
161
-
162
- Length 32 has the highest ego-agent success, ties for the highest
163
- cooperative-group success, and was the only setting whose three launches all
164
- completed below the evaluator's wiring guard.
 
 
 
 
 
 
 
 
 
 
 
165
 
166
  **SoccerTwos** β€” matching release-23 model-removed build, with the stock
167
  `SoccerTwos.onnx` from the same release as the opponent. Every row covers
 
76
  | `crawler` | `Tuple(Box(126), Box(32))` | `Box(20, [-1, 1])` |
77
  | `pushblock` | `Tuple(Box(105), Box(105))` | `Discrete(7)` |
78
 
79
+ The result column reports the published **ONNX at context length 32**. For
80
+ DungeonEscape, those ONNX files are retained legacy artifacts; the current
81
+ step-70,800 safetensors bundle and its cached-runtime results are documented
82
+ below. SoccerTwos ONNX remains derived from its sibling bundle.
 
83
 
84
  **Unity behavior names:** `DungeonEscape`, `SoccerTwos`. HF bundle subfolders are
85
  lowercase and case-sensitive.
 
89
  scene. `Batch` is the fixed ONNX batch: 36 for DungeonEscape (12 arenas Γ— 3
90
  agents), 16 for SoccerTwos (8 fields Γ— 2 controlled agents).
91
 
92
+ Training-data provenance is bundle-specific. The current DungeonEscape bundle
93
+ uses `simple-v0` + `medium-v0` + `expert-v0`; the SoccerTwos bundle remains
94
+ trained without expert trajectories.
95
 
96
  Metrics differ by scene type and are not comparable across rows. DungeonEscape is
97
  cooperative: **agent success** is the fraction of ego agents that finish with a
 
124
 
125
  | Bundle | Runtime | 16 (0.5Γ—) | 32 (1Γ—) | 64 (2Γ—) |
126
  |---|---|---:|---:|---:|
127
+ | `dungeon-escape` (legacy checkpoint) | PyTorch (cached) | 57.41% | **63.89%** | 61.11% |
128
+ | `dungeon-escape` (legacy ONNX) | ONNX (windowed) | 56.48% | **60.19%** | not published ‑ |
129
  | `soccer-twos` | PyTorch (cached) | β€” † | **41.25%** | β€” † |
130
  | `soccer-twos` | ONNX (windowed) | 40.00% | **40.00%** | 32.50% |
131
  | `pushblock` | PyTorch (cached) | 159/160 | 159/160 | 159/160 |
 
146
  <details>
147
  <summary><strong>Per-environment protocol and full results</strong></summary>
148
 
149
+ **DungeonEscape** β€” current PyTorch bundle, training step `70,800`,
150
+ matching release-23 model-removed build. This policy was trained with all three
151
+ data tiers: `simple-v0` + `medium-v0` + `expert-v0`.
 
152
 
153
+ Closed-loop cached-runtime evaluation used CUDA, 36 ego agents per launch
154
+ (12 cooperative groups of three), and Unity seeds `100..108`. Unity truncations
155
+ are reported as timeouts; episode length is not used to invalidate a seed.
156
+
157
+ | Runtime KV | Seeds | Agent success | Cooperative-group success | Natural termination |
158
+ |---:|---:|---:|---:|---:|
159
+ | 32 | `100..108` | 186/324 (57.41%) | **101/108 (93.52%)** | 322/324 |
160
+ | **64** | **`100..108`** | **187/324 (57.72%)** | **101/108 (93.52%)** | **324/324** |
161
+ | 128 (preliminary) | `100..102` | 63/108 (58.33%) | 35/36 (97.22%) | 107/108 |
162
+
163
+ KV64 is the recommended cached-runtime setting: it matches KV32 group success
164
+ while eliminating the two observed timeouts. The bundle's training context is
165
+ 32; KV64 and KV128 are inference-time retention settings.
166
+
167
+ The ONNX files currently in this folder are retained legacy deployment artifacts
168
+ from the preceding DungeonEscape checkpoint. Their published validation below
169
+ does **not** describe the step-70,800 safetensors bundle.
170
+
171
+ | Legacy ONNX context | Agent success | Cooperative-group success |
172
+ |---:|---:|---:|
173
+ | 16 | 61/108 (56.48%) | 33/36 (91.67%) |
174
+ | 32 | 65/108 (60.19%) | 34/36 (94.44%) |
175
 
176
  **SoccerTwos** β€” matching release-23 model-removed build, with the stock
177
  `SoccerTwos.onnx` from the same release as the opponent. Every row covers
dungeon-escape/config.json CHANGED
@@ -1,6 +1,7 @@
1
  {
2
  "bundle_format_version": 2,
3
- "package_version": "0.14.0",
 
4
  "model_config": {
5
  "network_name": "Llama",
6
  "num_layers": 4,
@@ -9,7 +10,7 @@
9
  "dropout": 0.05,
10
  "use_eos": true,
11
  "use_bos_action_gate": true,
12
- "rope_theta": 1000.0,
13
  "intermediate_size": 1024,
14
  "max_position_embeddings": 256,
15
  "context_length": 32
 
1
  {
2
  "bundle_format_version": 2,
3
+ "package": "causal-gpt-rl",
4
+ "package_version": "0.18.0",
5
  "model_config": {
6
  "network_name": "Llama",
7
  "num_layers": 4,
 
10
  "dropout": 0.05,
11
  "use_eos": true,
12
  "use_bos_action_gate": true,
13
+ "rope_theta": 10000.0,
14
  "intermediate_size": 1024,
15
  "max_position_embeddings": 256,
16
  "context_length": 32
dungeon-escape/metrics.json ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 70800,
3
+ "reasons": [
4
+ "best"
5
+ ],
6
+ "metrics": {
7
+ "eval_offline/action_nll": 1.7410476793746104,
8
+ "eval_offline/checkpoint_score": 0.46170878410339355
9
+ },
10
+ "metrics_step": 70800
11
+ }
dungeon-escape/model.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:e43c8a4f254c9b7a8090998e54885fad2328602d295bb6dcba9a62415751a182
3
  size 17194284
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d1e47d1139fd57f3af53d525586fe22a6924f99155e3825532814dae6b82ee12
3
  size 17194284