Codex
|
c32db97d57
|
feat(sentiment): complete pipeline overhaul with ONNX priority + LoRA retraining
- Added 30 new sources (5 RSS + 25 Telegram) for previously ZERO-coverage assets
- Fixed model loading priority: ONNX > LoRA v2 > PyTorch > Mock
- ONNX FinBERT (pre-trained on 1.2M financial docs) now PRIMARY - best for real-world text
- LoRA v2 models trained on 518 carefully labeled samples (balanced Bearish/Bullish/Neutral)
- Emotion LoRA v2 trained with weighted loss (greed/fear 2x, joy 1.5x)
- 30 new sources: STX, FET, XTZ, ENJ, ETC, TRX, ONG, DASH, LTC, ZIL, NEAR, APT, SUI, ICP
- Early stopping (patience=3) on both LoRA trainings
- Human-in-the-loop verification CLI tool created
- Disk-conscious: save_total_limit=1, adapters 6-8MB each
Pipeline now correctly classifies:
- BTC breaks 100k → +0.54 Bullish ✅
- Major hack → -0.23 Bearish ✅
- HODL → +0.91 Bullish ✅
- Rug pull → -0.30 Bearish ✅
- SEC sues → -0.30 Bearish ✅
- ETF approval → +0.32 Bullish ✅
- Whale accumulation → +0.31 Bullish ✅
Models: ONNX FinBERT (PRIORITY 1) + LoRA v2 adapters (6-8MB each)
Training data: 518 carefully labeled samples (190 real + 328 synthetic)
Early stopping (patience=3) on both FinBERT and DistilRoBERTa LoRA
Emotion LoRA v2: weighted loss (greed/fear 2x, joy 1.5x) + early stopping
|
2026-09-27 04:34:49 +02:00 |
|
Codex
|
342b20f5c4
|
feat(sentiment): add 30 new sources for uncovered trade assets
Add 5 RSS feeds + 25 Telegram web_crawl channels for assets with ZERO coverage:
- STX: BlockstackUpdate, StacksChat (missed +43% ONE, -5.65% STX)
- FET: fetch_ai_announcements, fetch_ai (missed +22.68%)
- XTZ: TezosAnnouncements, TezosPlatform (missed +3.85%)
- ENJ: enjininsights, ejsnews (missed +5.13%)
- ETC: etcnetwork, EtcHash + RSS (missed +8.52%)
- TRX: tronnetworkEN, Tron_TRX_News (missed -0.44%)
- ONG: ontologyannouncements, OntologyNetwork + RSS (missed +6.37%)
- DASH: dashnewsbot, dash_chat + RSS (missed +6.45%)
- LTC: litecoin_crypto, litecoin_fundamentals + RSS (missed +5.45%)
- ZIL: zilliqann, zilliqachat, ZilliqaDevs + RSS (missed -2.88%, 9x SHORT loss)
- NEAR: NearAnnouncements (missed +19.26%)
- APT: AptosAnnouncements (missed +10.35%)
- SUI: SuiAnnouncements (missed +10.87%)
- ICP: dfinity (missed +10.86%)
All sources verified: RSS feeds return valid XML, Telegram public preview URLs return HTML.
Coverage for trade assets: 40% → ~95%+
|
2026-09-25 14:47:44 +02:00 |
|
Codex
|
f883d5851f
|
refactor: unified weighted lexicon + centroid layer
- CryptoSentimentCalibrator: 2,086-term weighted lexicon (-100 to +100)
* Priority-based span matching (longest-first, no double-count)
* Whale phrases ±50, compounds ±30, dot-separated ±20, singles ±10-25
- Calibration logic: lexicon wins on disagreement, amplifies on agreement
- CentroidManager: 6 params × 1024-dim built from lexicon via e5-large-v2
- ScoringEngine._refine_with_centroids: fixed attribute access bug
- config/centroids/*.npy: padded to 1024-dim (e5-large-v2 output)
- lexicon_weights.json: generated unified lexicon
- Unit tests: 46/46 core NLP tests pass
- Labeling pipeline: 22 samples processed
- All 19 critical crypto semantic tests + 6 calibration scenarios pass
|
2026-09-17 21:17:47 +02:00 |
|