Model card: add dataset schema + example artifact + curriculum explanation
Browse files
README.md
CHANGED
|
@@ -28,6 +28,315 @@ This repository contains **PDT adapter/head weights** trained against the GPT-OS
|
|
| 28 |
Autoregressive decoding in Large Language Models (LLMs) is inherently sequential, creating a latency bottleneck that scales linearly with output length. While "Decomposition-and-Fill" methods like Skeleton-of-Thought attempt to parallelize generation via external orchestration, they suffer from coherence drift due to the lack of cross-stream communication. In this work, we introduce the Parallel Decoder Transformer (PDT), a parameter-efficient architecture that embeds coordination primitives directly into the inference process of a frozen pre-trained model. Instead of retraining the base model, PDT injects lightweight Speculative Note Conditioning (SNC) adapters that allow parallel decoding streams to synchronize via a shared, dynamic latent space. We formulate coordination as a speculative consensus problem, where sibling streams broadcast semantic "notes" to a global bus, gated by a learned verification head. We validate our approach on a 50,000-step curriculum using a frozen 20B-parameter backbone. Our results demonstrate that PDT achieves effective self-correction, reaching 77.8% precision in coverage prediction and recovering approximate serial semantics without modifying the trunk weights. This establishes PDT as a scalable, efficient alternative to full model fine-tuning for structured parallel generation.
|
| 29 |
|
| 30 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 31 |
## How to use
|
| 32 |
|
| 33 |
1. Install the reference implementation (runtime + scripts):
|
|
@@ -48,6 +357,25 @@ The complete training artifacts and dataset archives are mirrored publicly in GC
|
|
| 48 |
|
| 49 |
- **WandB run:** `https://wandb.ai/ljrweb-self/parallel-decoder-transformer/runs/fmuea63a`
|
| 50 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 51 |
## Citation
|
| 52 |
|
| 53 |
```bibtex
|
|
|
|
| 28 |
Autoregressive decoding in Large Language Models (LLMs) is inherently sequential, creating a latency bottleneck that scales linearly with output length. While "Decomposition-and-Fill" methods like Skeleton-of-Thought attempt to parallelize generation via external orchestration, they suffer from coherence drift due to the lack of cross-stream communication. In this work, we introduce the Parallel Decoder Transformer (PDT), a parameter-efficient architecture that embeds coordination primitives directly into the inference process of a frozen pre-trained model. Instead of retraining the base model, PDT injects lightweight Speculative Note Conditioning (SNC) adapters that allow parallel decoding streams to synchronize via a shared, dynamic latent space. We formulate coordination as a speculative consensus problem, where sibling streams broadcast semantic "notes" to a global bus, gated by a learned verification head. We validate our approach on a 50,000-step curriculum using a frozen 20B-parameter backbone. Our results demonstrate that PDT achieves effective self-correction, reaching 77.8% precision in coverage prediction and recovering approximate serial semantics without modifying the trunk weights. This establishes PDT as a scalable, efficient alternative to full model fine-tuning for structured parallel generation.
|
| 29 |
|
| 30 |
|
| 31 |
+
|
| 32 |
+
## Example: PDT notes artifact (truncated)
|
| 33 |
+
|
| 34 |
+
This is a real sample from the dataset pipeline (`survey_200141_ff0a0b4f.json`), shown with list/string truncation to keep the model card readable.
|
| 35 |
+
|
| 36 |
+
```json
|
| 37 |
+
{
|
| 38 |
+
"sample_id": "survey_200141_ff0a0b4f",
|
| 39 |
+
"domain": "survey",
|
| 40 |
+
"plan_path": "outputs/structured_plans/pdt_10k/survey/survey_200141_ff0a0b4f.json",
|
| 41 |
+
"sectional_independence": true,
|
| 42 |
+
"lag_delta": 1,
|
| 43 |
+
"note_cadence_M": 6,
|
| 44 |
+
"true_notes_example": {
|
| 45 |
+
"stream_id": "stream_1",
|
| 46 |
+
"ENT": [
|
| 47 |
+
{
|
| 48 |
+
"id": "E1",
|
| 49 |
+
"name": "Croatan",
|
| 50 |
+
"aliases": [
|
| 51 |
+
"Croatoan"
|
| 52 |
+
],
|
| 53 |
+
"type": "Ethnic Group",
|
| 54 |
+
"canonical": true
|
| 55 |
+
},
|
| 56 |
+
{
|
| 57 |
+
"id": "E2",
|
| 58 |
+
"name": "Dare County",
|
| 59 |
+
"aliases": [
|
| 60 |
+
"Alligator River",
|
| 61 |
+
"Croatan Sound",
|
| 62 |
+
"Roanoke Island",
|
| 63 |
+
"... <2 more items>"
|
| 64 |
+
],
|
| 65 |
+
"type": "Location",
|
| 66 |
+
"canonical": true
|
| 67 |
+
},
|
| 68 |
+
{
|
| 69 |
+
"id": "E3",
|
| 70 |
+
"name": "werowances",
|
| 71 |
+
"aliases": [
|
| 72 |
+
"chiefs"
|
| 73 |
+
],
|
| 74 |
+
"type": "Leadership Title",
|
| 75 |
+
"canonical": true
|
| 76 |
+
},
|
| 77 |
+
"... <4 more items>"
|
| 78 |
+
],
|
| 79 |
+
"FACT": [
|
| 80 |
+
{
|
| 81 |
+
"subj_id": "E1",
|
| 82 |
+
"predicate": "lived in",
|
| 83 |
+
"object": "coastal areas of what is now North Carolina",
|
| 84 |
+
"evidence_span": {
|
| 85 |
+
"start": 45,
|
| 86 |
+
"end": 87,
|
| 87 |
+
"text": "coastal areas of what is now North Carolina"
|
| 88 |
+
},
|
| 89 |
+
"certainty": 1.0
|
| 90 |
+
},
|
| 91 |
+
{
|
| 92 |
+
"subj_id": "E1",
|
| 93 |
+
"predicate": "might have been",
|
| 94 |
+
"object": "a branch of the larger Roanoke people or allied with them",
|
| 95 |
+
"evidence_span": {
|
| 96 |
+
"start": 92,
|
| 97 |
+
"end": 141,
|
| 98 |
+
"text": "a branch of the larger Roanoke people or allied with them"
|
| 99 |
+
},
|
| 100 |
+
"certainty": 0.8
|
| 101 |
+
},
|
| 102 |
+
{
|
| 103 |
+
"subj_id": "E2",
|
| 104 |
+
"predicate": "encompasses",
|
| 105 |
+
"object": "the Alligator River, Croatan Sound, Roanoke Island, Ocracoke Island, and parts of the Outer Banks",
|
| 106 |
+
"evidence_span": {
|
| 107 |
+
"start": 177,
|
| 108 |
+
"end": 265,
|
| 109 |
+
"text": "the Alligator River, Croatan Sound, Roanoke Island, Ocracoke Island, and parts of the Outer Banks"
|
| 110 |
+
},
|
| 111 |
+
"certainty": 1.0
|
| 112 |
+
},
|
| 113 |
+
"... <5 more items>"
|
| 114 |
+
],
|
| 115 |
+
"COVERAGE": [
|
| 116 |
+
{
|
| 117 |
+
"plan_item_id": "Define who the Croatan were, where they lived historically, and where related people live today.",
|
| 118 |
+
"status": "missing"
|
| 119 |
+
},
|
| 120 |
+
{
|
| 121 |
+
"plan_item_id": "Describe political leadership (werowances) and their responsibilities regarding wealth and decision-making.",
|
| 122 |
+
"status": "missing"
|
| 123 |
+
},
|
| 124 |
+
{
|
| 125 |
+
"plan_item_id": "Summarize core religious beliefs about a chief god, petty gods, immortality of the soul, heaven/Popogusso, and roles of priests and conjurors.",
|
| 126 |
+
"status": "missing"
|
| 127 |
+
},
|
| 128 |
+
"... <6 more items>"
|
| 129 |
+
]
|
| 130 |
+
},
|
| 131 |
+
"speculative_variant_example": {
|
| 132 |
+
"variant_id": "survey_200141_ff0a0b4f_variant_0",
|
| 133 |
+
"noise_config": {
|
| 134 |
+
"paraphrase_ratio": 0.15,
|
| 135 |
+
"drop_ratio": 0.05,
|
| 136 |
+
"hallucination_ratio": 0.05,
|
| 137 |
+
"shuffle_notes": true
|
| 138 |
+
},
|
| 139 |
+
"lag_delta": 1,
|
| 140 |
+
"notes_example": {
|
| 141 |
+
"stream_id": "stream_1",
|
| 142 |
+
"ENT": [
|
| 143 |
+
{
|
| 144 |
+
"id": "E1",
|
| 145 |
+
"name": "Croatan",
|
| 146 |
+
"aliases": [
|
| 147 |
+
"Croatoan",
|
| 148 |
+
"Croatian"
|
| 149 |
+
],
|
| 150 |
+
"type": "Ethnic Group",
|
| 151 |
+
"canonical": true
|
| 152 |
+
},
|
| 153 |
+
{
|
| 154 |
+
"id": "E2",
|
| 155 |
+
"name": "Dare County",
|
| 156 |
+
"aliases": [
|
| 157 |
+
"Alligator River",
|
| 158 |
+
"Croatan Sound",
|
| 159 |
+
"Roanoke Island",
|
| 160 |
+
"... <1 more items>"
|
| 161 |
+
],
|
| 162 |
+
"type": "Location",
|
| 163 |
+
"canonical": true
|
| 164 |
+
},
|
| 165 |
+
{
|
| 166 |
+
"id": "E3",
|
| 167 |
+
"name": "werowances",
|
| 168 |
+
"aliases": [
|
| 169 |
+
"chiefs",
|
| 170 |
+
"leaders"
|
| 171 |
+
],
|
| 172 |
+
"type": "Leadership Title",
|
| 173 |
+
"canonical": true
|
| 174 |
+
},
|
| 175 |
+
"... <1 more items>"
|
| 176 |
+
],
|
| 177 |
+
"FACT": [
|
| 178 |
+
{
|
| 179 |
+
"subj_id": "E1",
|
| 180 |
+
"predicate": "lived in",
|
| 181 |
+
"object": "coastal areas of what is now North Carolina",
|
| 182 |
+
"evidence_span": {
|
| 183 |
+
"start": 45,
|
| 184 |
+
"end": 87,
|
| 185 |
+
"text": "coastal areas of what is now North Carolina"
|
| 186 |
+
},
|
| 187 |
+
"certainty": 1.0
|
| 188 |
+
},
|
| 189 |
+
{
|
| 190 |
+
"subj_id": "E1",
|
| 191 |
+
"predicate": "might have been",
|
| 192 |
+
"object": "a branch of the larger Roanoke people or allied with them",
|
| 193 |
+
"evidence_span": {
|
| 194 |
+
"start": 92,
|
| 195 |
+
"end": 141,
|
| 196 |
+
"text": "a branch of the larger Roanoke people or allied with them"
|
| 197 |
+
},
|
| 198 |
+
"certainty": 0.8
|
| 199 |
+
},
|
| 200 |
+
{
|
| 201 |
+
"subj_id": "E2",
|
| 202 |
+
"predicate": "encompasses",
|
| 203 |
+
"object": "the Alligator River, Croatan Sound, Roanoke Island, Ocracoke Island, and parts of the Outer Banks",
|
| 204 |
+
"evidence_span": {
|
| 205 |
+
"start": 177,
|
| 206 |
+
"end": 265,
|
| 207 |
+
"text": "the Alligator River, Croatan Sound, Roanoke Island, Ocracoke Island, and parts of the Outer Banks"
|
| 208 |
+
},
|
| 209 |
+
"certainty": 1.0
|
| 210 |
+
},
|
| 211 |
+
"... <1 more items>"
|
| 212 |
+
],
|
| 213 |
+
"COVERAGE": [
|
| 214 |
+
{
|
| 215 |
+
"plan_item_id": "Define who the Croatan were, where they lived historically, and where related people live today.",
|
| 216 |
+
"status": "missing"
|
| 217 |
+
},
|
| 218 |
+
{
|
| 219 |
+
"plan_item_id": "Describe political leadership (werowances) and their responsibilities regarding wealth and decision-making.",
|
| 220 |
+
"status": "missing"
|
| 221 |
+
},
|
| 222 |
+
{
|
| 223 |
+
"plan_item_id": "Summarize core religious beliefs about a chief god, petty gods, immortality of the soul, heaven/Popogusso, and roles of priests and conjurors.",
|
| 224 |
+
"status": "missing"
|
| 225 |
+
},
|
| 226 |
+
"... <1 more items>"
|
| 227 |
+
]
|
| 228 |
+
}
|
| 229 |
+
},
|
| 230 |
+
"versioned_notes_snapshot_0": {
|
| 231 |
+
"snapshot_id": 0,
|
| 232 |
+
"source": "procedural_bus",
|
| 233 |
+
"lag_delta": 1,
|
| 234 |
+
"note_cadence_M": 6,
|
| 235 |
+
"ent_count": 9,
|
| 236 |
+
"fact_count": 10,
|
| 237 |
+
"notes_example": {
|
| 238 |
+
"stream_id": "stream_1",
|
| 239 |
+
"ENT": [
|
| 240 |
+
{
|
| 241 |
+
"id": "E1",
|
| 242 |
+
"name": "Croatan",
|
| 243 |
+
"aliases": [
|
| 244 |
+
"Croatoan"
|
| 245 |
+
],
|
| 246 |
+
"type": "Ethnic Group",
|
| 247 |
+
"canonical": true
|
| 248 |
+
},
|
| 249 |
+
{
|
| 250 |
+
"id": "E2",
|
| 251 |
+
"name": "Dare County",
|
| 252 |
+
"aliases": [
|
| 253 |
+
"Alligator River",
|
| 254 |
+
"Croatan Sound",
|
| 255 |
+
"Roanoke Island",
|
| 256 |
+
"... <1 more items>"
|
| 257 |
+
],
|
| 258 |
+
"type": "Location",
|
| 259 |
+
"canonical": true
|
| 260 |
+
},
|
| 261 |
+
{
|
| 262 |
+
"id": "E3",
|
| 263 |
+
"name": "werowances",
|
| 264 |
+
"aliases": [
|
| 265 |
+
"chiefs"
|
| 266 |
+
],
|
| 267 |
+
"type": "Leadership Title",
|
| 268 |
+
"canonical": true
|
| 269 |
+
},
|
| 270 |
+
"... <1 more items>"
|
| 271 |
+
],
|
| 272 |
+
"FACT": [
|
| 273 |
+
{
|
| 274 |
+
"subj_id": "E1",
|
| 275 |
+
"predicate": "lived in",
|
| 276 |
+
"object": "coastal areas of what is now North Carolina",
|
| 277 |
+
"evidence_span": {
|
| 278 |
+
"start": 45,
|
| 279 |
+
"end": 87,
|
| 280 |
+
"text": "coastal areas of what is now North Carolina"
|
| 281 |
+
},
|
| 282 |
+
"certainty": 1.0
|
| 283 |
+
},
|
| 284 |
+
{
|
| 285 |
+
"subj_id": "E1",
|
| 286 |
+
"predicate": "might have been",
|
| 287 |
+
"object": "a branch of the larger Roanoke people or allied with them",
|
| 288 |
+
"evidence_span": {
|
| 289 |
+
"start": 92,
|
| 290 |
+
"end": 141,
|
| 291 |
+
"text": "a branch of the larger Roanoke people or allied with them"
|
| 292 |
+
},
|
| 293 |
+
"certainty": 0.8
|
| 294 |
+
},
|
| 295 |
+
{
|
| 296 |
+
"subj_id": "E2",
|
| 297 |
+
"predicate": "encompasses",
|
| 298 |
+
"object": "the Alligator River, Croatan Sound, Roanoke Island, Ocracoke Island, and parts of the Outer Banks",
|
| 299 |
+
"evidence_span": {
|
| 300 |
+
"start": 177,
|
| 301 |
+
"end": 265,
|
| 302 |
+
"text": "the Alligator River, Croatan Sound, Roanoke Island, Ocracoke Island, and parts of the Outer Banks"
|
| 303 |
+
},
|
| 304 |
+
"certainty": 1.0
|
| 305 |
+
},
|
| 306 |
+
"... <1 more items>"
|
| 307 |
+
],
|
| 308 |
+
"COVERAGE": [
|
| 309 |
+
{
|
| 310 |
+
"plan_item_id": "Define who the Croatan were, where they lived historically, and where related people live today.",
|
| 311 |
+
"status": "missing"
|
| 312 |
+
},
|
| 313 |
+
{
|
| 314 |
+
"plan_item_id": "Describe political leadership (werowances) and their responsibilities regarding wealth and decision-making.",
|
| 315 |
+
"status": "missing"
|
| 316 |
+
},
|
| 317 |
+
{
|
| 318 |
+
"plan_item_id": "Summarize core religious beliefs about a chief god, petty gods, immortality of the soul, heaven/Popogusso, and roles of priests and conjurors.",
|
| 319 |
+
"status": "missing"
|
| 320 |
+
},
|
| 321 |
+
"... <1 more items>"
|
| 322 |
+
]
|
| 323 |
+
}
|
| 324 |
+
},
|
| 325 |
+
"rollback": {
|
| 326 |
+
"triggered": false,
|
| 327 |
+
"l_tokens": 0,
|
| 328 |
+
"events": []
|
| 329 |
+
}
|
| 330 |
+
}
|
| 331 |
+
```
|
| 332 |
+
|
| 333 |
+
To reproduce this view locally:
|
| 334 |
+
|
| 335 |
+
```bash
|
| 336 |
+
uv run python scripts/pretty_notes_artifact.py survey_200141_ff0a0b4f.json
|
| 337 |
+
```
|
| 338 |
+
|
| 339 |
+
|
| 340 |
## How to use
|
| 341 |
|
| 342 |
1. Install the reference implementation (runtime + scripts):
|
|
|
|
| 357 |
|
| 358 |
- **WandB run:** `https://wandb.ai/ljrweb-self/parallel-decoder-transformer/runs/fmuea63a`
|
| 359 |
|
| 360 |
+
## Why the dataset is structured this way
|
| 361 |
+
|
| 362 |
+
PDT is trained on **streamed, structured supervision** produced by a 5-stage pipeline:
|
| 363 |
+
|
| 364 |
+
- **Stage 2 (Plans):** a 3-stream decomposition plan is generated for each document.
|
| 365 |
+
- **Stage 3 (Notes):** we generate **true notes (teacher)** and **speculative notes (student input)** in a consistent schema:
|
| 366 |
+
- `ENT`: entity table (stable ids)
|
| 367 |
+
- `FACT`: grounded tuples with `evidence_span`
|
| 368 |
+
- `COVERAGE`: plan-item status targets (`covered|partial|missing`)
|
| 369 |
+
- `versioned_notes`: lagged, versioned snapshots mirroring the Dynamic Notes Bus semantics
|
| 370 |
+
- **Stage 5 (KD Export):** these artifacts are converted into `kd_*.jsonl` where each line is a **stream-level** training example.
|
| 371 |
+
|
| 372 |
+
This layout is required to support the **teacher→student curriculum** described in the training guide:
|
| 373 |
+
|
| 374 |
+
- **Stage 0:** planner/notes-head bootstrap (trunk frozen)
|
| 375 |
+
- **Stage 1:** stream adapters + SNC cross-attention bootstrap (speculation frozen; teacher notes forced)
|
| 376 |
+
- **Stage 2:** enable speculation + notes-bus usage (teacher-heavy mixing)
|
| 377 |
+
- **Stage 3:** train agreement + coverage heads for self-correction/rollback behavior (still trunk frozen)
|
| 378 |
+
|
| 379 |
## Citation
|
| 380 |
|
| 381 |
```bibtex
|