Add v1 data description for raw/direct/profilegen/retrievalgen support set
Browse files- v1_DATA_DESCRIPTION.md +77 -0
v1_DATA_DESCRIPTION.md
ADDED
|
@@ -0,0 +1,77 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# UserModeling Dataset — v1 Data Description
|
| 2 |
+
|
| 3 |
+
This repository contains raw data and three benchmark tracks for user modeling on market stance prediction.
|
| 4 |
+
|
| 5 |
+
## 1) Raw source
|
| 6 |
+
|
| 7 |
+
- `raw/data_v1.parquet`
|
| 8 |
+
- Canonical raw source file for v1.
|
| 9 |
+
- Includes wallet-level market history with block-range fields.
|
| 10 |
+
|
| 11 |
+
## 2) v1 benchmark tracks
|
| 12 |
+
|
| 13 |
+
### A. Direct track
|
| 14 |
+
- Path: `v1_direct/benchmark_core_direct/`
|
| 15 |
+
- Files:
|
| 16 |
+
- `train.parquet`
|
| 17 |
+
- `validation.parquet`
|
| 18 |
+
- `test.parquet`
|
| 19 |
+
- Purpose:
|
| 20 |
+
- Single-step prediction from user history to target market.
|
| 21 |
+
|
| 22 |
+
### B. ProfileGen (LLM) track
|
| 23 |
+
- Path: `v1_profilegen/benchmark_profilegen/`
|
| 24 |
+
- Files:
|
| 25 |
+
- `train.parquet`
|
| 26 |
+
- `validation.parquet`
|
| 27 |
+
- `test.parquet`
|
| 28 |
+
- Purpose:
|
| 29 |
+
- Two-step process:
|
| 30 |
+
1) generate user profile with LLM
|
| 31 |
+
2) predict target market choice/confidence using generated profile
|
| 32 |
+
|
| 33 |
+
### C. RetrievalGen support set
|
| 34 |
+
- Path: `v1_retrievalgen/support_set/`
|
| 35 |
+
- Files:
|
| 36 |
+
- `support.parquet`
|
| 37 |
+
- `wallet_ids_support.txt`
|
| 38 |
+
- `support_spec.md`
|
| 39 |
+
- `support_manifest.json`
|
| 40 |
+
- Purpose:
|
| 41 |
+
- Support pool for retrieval-based two-step methods.
|
| 42 |
+
- Constructed by excluding the 2000 experiment wallets.
|
| 43 |
+
|
| 44 |
+
## 3) Target task definition (shared by Direct/ProfileGen/RetrievalGen)
|
| 45 |
+
|
| 46 |
+
For each target market sample, predict:
|
| 47 |
+
- `choice`
|
| 48 |
+
- `confidence`
|
| 49 |
+
|
| 50 |
+
So DirectGen and RetrievalGen share the same end task, with different input pipelines.
|
| 51 |
+
|
| 52 |
+
## 4) Time / block fields
|
| 53 |
+
|
| 54 |
+
Block-range fields are preserved in v1 assets:
|
| 55 |
+
- `stance_first_block`
|
| 56 |
+
- `stance_last_block`
|
| 57 |
+
|
| 58 |
+
These fields can be used to enforce time-aware filtering (e.g., no-future retrieval).
|
| 59 |
+
|
| 60 |
+
## 5) Split and wallet metadata
|
| 61 |
+
|
| 62 |
+
See split files under each track (`splits/`) for exact split config and wallet lists.
|
| 63 |
+
|
| 64 |
+
## 6) Minimal loading example (Python)
|
| 65 |
+
|
| 66 |
+
```python
|
| 67 |
+
import pyarrow.parquet as pq
|
| 68 |
+
|
| 69 |
+
train = pq.read_table('v1_direct/benchmark_core_direct/train.parquet')
|
| 70 |
+
print(train.num_rows, train.column_names)
|
| 71 |
+
|
| 72 |
+
profile_test = pq.read_table('v1_profilegen/benchmark_profilegen/test.parquet')
|
| 73 |
+
print(profile_test.num_rows)
|
| 74 |
+
|
| 75 |
+
support = pq.read_table('v1_retrievalgen/support_set/support.parquet')
|
| 76 |
+
print(support.num_rows)
|
| 77 |
+
```
|