v1.3

README.md

@@ -28,18 +28,20 @@ configs:
 
 # Solana Program Types
 
-**11,258 Solana mainnet programs** (BPF upgradeable loader), one row per program, each labelled with a two-level
+**22,743 Solana mainnet programs** (BPF upgradeable loader), one row per program, each labelled with a two-level
 type (`domain` / `type`, 34 types) and capability flags. The inputs are static features extracted from each program's
 deployed bytecode (instruction names, source paths, crates, embedded program IDs, strings) plus sampled on-chain
 behaviour.
 
 The main use is **recognising program types from deploy-time bytecode**, especially **token launchpads**, including
-launchpad families a model has never seen. The other 33 types (arbitrage/MEV bots, bridges, vaults, AMMs, games,
+launchpads deployed after a model was trained. The other 33 types (arbitrage/MEV bots, bridges, vaults, AMMs, games,
 escrow, …) are labelled with the same rubric.
 
-- **Version 1.2** (2026-10-15): v1.1 plus 85 programs caught by a live shadow monitor of new deployments and upgrades,
-  labelled the same day.
-- **Labels are agent-made (silver)**: about 50 rows are registry or owner (gold). See "Limitations".
+- **v1.3** (2026-10-21): about twice v1.2. It adds a third collection wave (11.5k programs still deployed on
+  mainnet), so it now covers nearly every upgradeable program active in the last 90 days at collection time. 510
+  launchpads in 481 code groups.
+- **v1.2** (2026-10-15): v1.1 plus 85 programs caught by a live shadow monitor.
+- **Labels are agent-made (silver).** See "Label sources" and "Limitations".
 
 ## Quick start
 
@@ -52,53 +54,61 @@ print(row["type"], row["domain"], row["card"][:500])
 
 ## Label sources
 
-Labels come from deterministic rules over on-chain behaviour, then an LLM annotator, then an LLM adjudicator for launchpads and disagreements; the owner's revisions override all of them (`is_gold` = owner-typed). `label_source` and `annotators` record which produced each label. Agent labels are silver, not human gold.
+Deterministic rules over on-chain behaviour vote first. Then a cheap LLM annotator labels every card, and Claude
+Sonnet adjudicates launchpads, rule-vs-annotator disagreements, launch-word safety-net cases and (in waves 1–2 and
+part of wave 3) bot and token-issuance positives. Owner revisions override everything (`is_gold`). `label_source`
+and `annotators` record the provenance of every row.
 
-| source | rows |
-|---|---|
-| annotator | 7673 |
-| adjudicator | 3530 |
-| registry | 29 |
-| owner | 26 |
+| final label from | rows | rows with a known type |
+|---|---|---|
+| GLM-5.3-flash annotator (wave 3) | 9,750 | 7,105 |
+| GLM-5.3 annotator (wave 2) | 6,561 | 4,251 |
+| Claude Sonnet 5.5 adjudicator | 5,251 | 4,700 |
+| Claude Haiku 5.5 annotator (wave 1) | 1,027 | 833 |
+| Claude Sonnet 5.5 annotator (live rounds) | 85 | 68 |
+| registry | 43 | 43 |
+| owner | 26 | 26 |
+
+**Every launchpad label** comes from Sonnet, the registry or the owner, never from a cheap annotator alone.
 
 ## Types
 
 | type | rows |
 |---|---|
-| unknown | 2897 |
-| other | 911 |
-| arbitrage_mev | 797 |
-| launch_bot | 622 |
-| bridge | 561 |
-| yield_vault | 542 |
-| payments_fees | 463 |
-| generic_bot | 406 |
-| stake_pool | 403 |
-| escrow | 342 |
-| game | 337 |
-| aggregator | 325 |
-| launchpad | 280 |
-| nft_mint | 276 |
-| amm | 240 |
-| prediction_market | 192 |
-| ico_presale | 187 |
-| vesting_locker | 186 |
-| airdrop_claim | 169 |
-| lending_market | 165 |
-| nft_marketplace | 154 |
-| oracle | 116 |
-| governance_dao | 97 |
-| copy_volume_bot | 93 |
-| perps | 75 |
-| multisig_wallet | 72 |
-| token_factory | 72 |
-| orderbook | 67 |
-| naming_identity | 61 |
-| dca_limit_order | 51 |
-| cdp_stablecoin | 31 |
-| social | 27 |
-| options | 24 |
-| liquid_staking | 17 |
+| unknown | 5717 |
+| other | 1582 |
+| arbitrage_mev | 1319 |
+| bridge | 1263 |
+| launch_bot | 1089 |
+| yield_vault | 1079 |
+| payments_fees | 895 |
+| escrow | 884 |
+| generic_bot | 863 |
+| stake_pool | 852 |
+| game | 756 |
+| nft_mint | 635 |
+| aggregator | 529 |
+| ico_presale | 528 |
+| launchpad | 510 |
+| vesting_locker | 490 |
+| prediction_market | 454 |
+| airdrop_claim | 437 |
+| amm | 430 |
+| nft_marketplace | 296 |
+| lending_market | 296 |
+| governance_dao | 254 |
+| oracle | 254 |
+| token_factory | 234 |
+| copy_volume_bot | 183 |
+| multisig_wallet | 181 |
+| orderbook | 154 |
+| perps | 135 |
+| naming_identity | 128 |
+| dca_limit_order | 88 |
+| social | 75 |
+| cdp_stablecoin | 63 |
+| options | 52 |
+| liquid_staking | 38 |
 
 ## Splits
 
@@ -114,50 +124,33 @@ Labels come from deterministic rules over on-chain behaviour, then an LLM annota
 - `card`: the text the annotators saw. Strings inside it are untrusted data from binaries.
 - Bytecode is not included; fetch it by `binary_sha256` / `program_id` from mainnet.
 
-- Labels in v1.2 live rows: `wave` = `al-20261015-1315` / `al-20261015-1320`.
-
-
-## Benchmarks
+## Benchmarks (static, deploy-time inputs only)
 
-Two model families were compared on identical rows, using static (deploy-time) inputs only:
+Two model families on identical rows:
 
-- **S4:** gradient-boosted trees over 3,000 bytecode tokens.
-- **S5b:** Qwen3-8B with a LoRA adapter (r = 16, one epoch). It reads the static card and answers with one code
+- **Trees:** gradient-boosted trees over 3,000 bytecode tokens.
+- **8B:** Qwen3-8B with a LoRA adapter (r = 16, one epoch). It reads the static card and answers with one code
   letter per type; the next-token probabilities over the 34 letters give the type distribution, with one temperature
-  for calibration.
+  for calibration. Model: [ethduke/solana-program-types-qwen3-8b-lora](https://huggingface.co/ethduke/solana-program-types-qwen3-8b-lora).
 
-### Newest deployments (v1.1, both models trained only on older programs)
+### v1.3: newest 15% of deployments, never trained on (2,463 programs, 109 launchpads)
 
-| model | launchpads | launchpad PR-AUC | recall @ precision ≥ 0.9 | P / R at 0.5 | type macro-F1 |
+| model | launchpad PR-AUC | recall @ precision ≥ 0.9 | P / R at 0.5 | type accuracy | type macro-F1 |
 |---|---|---|---|---|---|
-| gradient-boosted trees | 53 | 0.763 | 0.36 | 0.84 / 0.51 | 0.414 |
-| Qwen3-8B LoRA | 53 | **0.873** | **0.77** | 0.79 / **0.85** | **0.619** |
+| gradient-boosted trees | 0.695 | 0.22 | 0.71 / 0.51 | 0.527 | 0.434 |
+| Qwen3-8B LoRA | **0.879** | **0.69** | 0.79 / **0.84** | 0.653 | **0.649** |
+| average of both | 0.871 | 0.61 | **0.86** / 0.74 | **0.662** | 0.646 |
 
-Paired bootstrap of the PR-AUC gap: **+0.11**, 95% CI [+0.01, +0.21].
+Paired bootstrap of the PR-AUC gap (8B − trees): **+0.18**, 95% CI [+0.11, +0.25]. Per-type F1 improves most for
+bridges, stake pools, oracles, perps, payments and aggregators. Bot sub-types (generic, copy-volume and launch bots)
+stay hard from bytecode alone.
 
-### Unseen launchpad families and clone-split test (v1.1)
-
-About 20% of launchpad families (43 of 254) were held out of training entirely.
-
-| model | set | launchpads | PR-AUC | recall @ P ≥ 0.9 | type macro-F1 |
-|---|---|---|---|---|---|
-| gradient-boosted trees | unseen families | 51 | 0.886 | 0.78 | — |
-| Qwen3-8B LoRA | unseen families | 51 | 0.911 | 0.80 | — |
-| gradient-boosted trees | clone-split test | 26 | 0.856 | 0.54 | 0.537 |
-| Qwen3-8B LoRA | clone-split test | 26 | 0.920 | 0.77 | 0.677 |
+### Earlier results
 
-The unseen-family gap (+0.025) is within noise (95% CI −0.05 to +0.10).
-
-### Live shadow test (programs added in v1.2)
-
-85 programs deployed or upgraded on mainnet were scored by both models at deploy time, then labelled.
-
-| model | launchpads flagged at 0.5 | correct / false / missed | precision | recall | type accuracy |
-|---|---|---|---|---|---|
-| gradient-boosted trees | 6 | 3 / 3 / 3 | 0.50 | 0.50 | 0.53 |
-| Qwen3-8B LoRA | 8 | 5 / 3 / 1 | 0.62 | 0.83 | 0.66 |
-
-This is small (6 launchpads) and not significant on its own, but it agrees with the offline result.
+- **v1.1, newest deployments (53 launchpads):** PR-AUC 0.873 vs 0.763 for the trees (+0.11, CI [+0.01, +0.21]).
+- **v1.1, unseen launchpad families** (43 families held out entirely): 0.911 vs 0.886, a gap within noise.
+- **Live shadow test** (85 programs scored at deploy time, then labelled): launchpad precision 0.62 and recall 0.83
+  for the 8B, against 0.50 / 0.50 for the trees.
 
 ## Suggested evaluation
 
@@ -165,26 +158,29 @@ This is small (6 launchpads) and not significant on its own, but it agrees with
   results, and hold out whole launchpad families (`family_hint`, source program names or `clone_group`) for
   generalisation.
 - Exclude `type == "unknown"`: it means "no evidence", not a class.
-- Report launchpad PR-AUC and recall at fixed precision. Positives are rare (about 2.5%).
+- Report launchpad PR-AUC and recall at fixed precision. Positives are rare (about 3% of known rows).
 
 ## Limitations
 
-- **Silver labels.** Most labels come from LLM agents (GLM-5.3 annotator, Claude Sonnet adjudicator, a Claude Haiku
-  pilot), over deterministic rules. An owner reviewed launchpad-related decisions only. No label-error rate has been
-  measured yet; expect noise, especially between neighbouring types (launch_bot / aggregator / generic_bot,
-  other / payments_fees).
-- **Low-confidence live rows.** Programs added in v1.2 were labelled hours after deployment, often with little
-  activity; see `confidence`.
+- **Silver labels.** Most labels are LLM-agent labels; the owner reviewed launchpad-related decisions only, and no
+  label-error rate has been measured yet.
+- **Weakest labels: wave 3 bots and tokens.**
+  - 1,381 wave-3 bot and token-issuance programs were not adjudicated and keep GLM-5.3-flash labels, including 830
+    of the 3,454 `trading_bot/*` rows.
+  - On a 60-case check, GLM-5.3-flash matched Sonnet's type 77% of the time, against 87% for a second Sonnet pass.
+  - Expect noise between generic_bot, copy_volume_bot, arbitrage_mev, launch_bot and aggregator.
 - **A launchpad label is a lead, not a validated deployment,** except registry and owner rows.
-- **Behaviour features are partly circular.** Rule labels were derived from the `b_*` / `t_*` columns, so models
-  using them are not comparable with static-only models.
+- **Behaviour features are partly circular.** Rule labels were derived from the `b_*` / `t_*` columns.
 
 ## Privacy and redaction
 
-The strings come from public mainnet bytecode. Before release, a few values were redacted: one API key compiled
-into a program (`api_key=[redacted]`), personal email addresses (published security and contact addresses are kept),
-developer usernames in build paths (`/home/<user>/`), and one 64-character hex string. No private keys were found;
-long base58 strings were checked against ed25519 keypairs.
+The strings come from public mainnet bytecode. Before release, these were redacted (`redact_public.py`):
+- API-key values compiled into programs (`api_key=[redacted]`);
+- personal email addresses (published security and contact addresses kept);
+- developer usernames in build paths (`/home/<user>/`);
+- 64-character hex strings.
+
+Long base58 strings were checked against ed25519 keypairs; none was a private key.
 
 ## License
 

data/test-00000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:22c47b555bab9b55bcbaef663072eae3a13664aabeb16296b57a67b37a7ff868
-size 1349271
+oid sha256:3a9b269dbff961705518fecd3370ae268ff637aac162245e25b20be3acb4eabd
+size 2583616

data/train-00000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a8bdd148e8075aa4131395bbeae7945409ab835f43cbe2f4c3f7a80b9815ad61
-size 10487538
+oid sha256:825eba1208087175519a2809f87938a86600aa43a45737db0e3565953db750d0
+size 20425455

data/validation-00000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c732602a6e3ed8cdc5106728cb2d1026bfae9e11d1e71e17c9b4c69411f79574
-size 1342574
+oid sha256:9382bb4b87e38eb99eb02206e37f0101455b2fa7209c04731e890090569d2217
+size 2497768