SELI Auditor BF16

SELI Solidity audit artifact.

Variant

Merged BF16 Hugging Face model. This variant contains the seed adapter plus the local SELI fine-tune merged into the Qwen2.5-Coder base.

Training stack:

Qwen/Qwen2.5-Coder-7B-Instruct
  -> merge seed adapter: jhsu12/solidity-vulnerability-detector
  -> train local SELI LoRA on audit dataset
  -> merge local SELI LoRA
  -> artifact: BF16 safetensors model

Use this variant for highest quality Transformers/vLLM/SGLang serving when VRAM allows.

Base And Fine-Tune Stack

  • Base model: Qwen/Qwen2.5-Coder-7B-Instruct
  • HF card base model: Qwen/Qwen2.5-Coder-7B-Instruct
  • Seed adapter: jhsu12/solidity-vulnerability-detector
  • Local fine-tune: SELI audit dataset
  • Artifact: merged BF16 model
  • Created: 2026-05-14 01:52:51 UTC

Training Data Coverage

Counts below are computed from the context-filtered Qwen2.5-Coder training files available during upload.

Metric Value
Raw JSONL rows 4072
Context-filtered rows used 3441
Rows dropped for context overflow 631
Max training context 4096 tokens
Named bug shapes 95
Shape mentions 570
Coverage axes 111
Proof gates 69
Bypass-hypothesis rows 570
Line-label rows 570

Strong Shape Families

  • solidity: 3204
  • access-control: 3106
  • amm: 3095
  • accounting: 2571
  • decompiler: 2504
  • reentrancy: 2481
  • oracle: 2299
  • upgradeability: 2112
  • foundry: 1967
  • money-flow: 1832
  • flashloan: 1786
  • economic-model: 1664
  • signature: 1554
  • defi-hacks: 1383
  • constructor-plane: 1187
  • fork-poc: 1168
  • incident-reconstruction: 1168
  • multi-tx-window: 1113
  • ast-ir: 700
  • proof-gates: 591
  • line-level: 570
  • regression: 570
  • tool-use: 562
  • bridge-trust-config: 541

Source Families

  • historical_poc_foundry: 789
  • defihacklabs_incident_index: 594
  • curated_line_level_regression_suite: 570
  • local_skill_archive: 499
  • static_tool_artifact: 496
  • local_skill: 192
  • curated_vulnerable_contracts: 143
  • foundry_vulnerability_labs: 54
  • taxonomy_and_testcases: 36
  • ctf_training: 23
  • incident_intel_feed: 4
  • curated_regression: 4
  • audit_finding_dataset: 3
  • contest_judging_guidelines: 2
  • decompiler_tooling: 2
  • security_standard: 2
  • incident_database: 2
  • audit_model_adapter_index: 1
  • static_analyzer_json_schema: 1
  • curated_vulnerability_dataset: 1

Named Shape Coverage

Shape Count
epoch snapshot reward shared-pool reentrancy 6
read-only reentrancy shared-pool price cache 6
entrypoint asymmetry 6
shared-escrow partial claim clear 6
delegated operator shared-pool reentrancy 6
emergency-withdraw shared-pool accounting bypass 6
cross-domain auth 6
shared ETH refund nonce callback 6
nested multicall deferred settlement shared-pool reentrancy 6
stale liability cache shared-pool accounting 6
liquidation callback shared collateral 6
stale oracle 6
async bridge finalize shared-escrow reentrancy 6
arbitrary call allowance drain 6
whitelist trycatch token callback shared-pool reentrancy 6
Foundry invariant proof 6
ZK public input binding 6
Merkle replay 6
multicall deferred-settlement shared-pool reentrancy 6
approval callback claim reuse shared-pool reentrancy 6
signed delta cast 6
referral MLM shared reward-pool reentrancy 6
Heimdall decompiled CFG 6
ERC4626 share inflation 6
router liquidity delta shared-pool reentrancy 6
AST IR state mapping 6
governance flashloan 6
proxy storage collision 6
read-only reentrancy 6
ERC1155 receiver shared-pool reentrancy 6
cross-contract guard-domain shared-pool reentrancy 6
router balance-sweep shared-pool reentrancy 6
rescue sweep alias shared-pool bypass 6
shared-pool reentrancy two-user harness 6
static analyzer false negative 6
selector ambiguity 6
multi-asset basket partial-settle shared-pool reentrancy 6
EIP-7702 delegated EOA bypass 6
cross-function reentrancy 6
cooldown alias shared-pool reentrancy 6
liquidation shared collateral reentrancy 6
queued-claim reentrancy delete-after-transfer 6
NFT callback reentrancy 6
slippage/mep 6
nominal LP accounting shared-pool bypass 6

Proof-Gated Audit Behaviors

  • SOURCE_AVAILABILITY_CHECK: 570
  • MANUAL_REVIEW_PACKET: 570
  • STATE_WRITER_READER_TABLE: 570
  • FALSE_POSITIVE_CONTROLS: 570
  • MONEY_FLOW_GRAPH: 396
  • PATH_PROFIT_SWEEP: 396
  • CALL_FLOW_TRACE: 330
  • TWO_USER_SHARED_POOL_POC: 288
  • SUM_CLAIMS_LTE_ASSETS_INVARIANT: 288
  • FOUNDRY_POC: 144
  • FOUNDRY_TRACE: 144
  • BRANCH_BYPASS_MATRIX: 120
  • RECEIVER_HOOK_TRACE: 48
  • TOKENIZED_POSITION_CLAIM_BURN_CHECK: 48
  • BALANCE_DELTA_PARITY_CHECK: 42
  • ALLOWANCE_SIDE_EFFECT_TRACE: 30
  • CLAIM_CONSUMPTION_BEFORE_APPROVAL_CHECK: 30
  • LOCK_SCOPE_MATRIX: 30
  • MULTI_ENTRYPOINT_REENTRANCY_POC: 30
  • LIQUIDATION_COMMIT_ORDER_CHECK: 24
  • BORROWER_DEBT_COLLATERAL_INVARIANT: 24
  • PER_USER_CLAIM_CLOSE_BEFORE_CALLBACK_CHECK: 24
  • PREDICATE_SPECIFIC_CEI_TRACE: 24
  • MESSAGE_CONSUMPTION_BEFORE_PAYOUT_CHECK: 18

Coverage Axes

  • line_level_evidence: 570
  • invariant_binding: 570
  • near_miss_control: 570
  • asset_liability_delta: 378
  • shared_pool_claims: 378
  • rounding_and_donation: 378
  • two_user_harness: 288
  • victim_backing: 288
  • sum_claims_lte_assets: 288
  • per_user_claim_invalidation: 288
  • self_reentry: 276
  • cross_function_reentry: 276
  • post_callback_state_write: 276
  • authority_source: 120
  • downstream_identity: 120
  • nonce_domain_binding: 120
  • receiver_hook_reachability: 48
  • tokenized_claim_burn_order: 48
  • batch_id_reordering: 48
  • staleness: 42
  • same_tx_manipulation: 42
  • liquidity_depth: 42
  • source_domain: 30
  • message_nonce: 30

Minimal Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "0xtoshi/seli_auditor-BF16"
tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="bfloat16",
    device_map="auto",
    trust_remote_code=True,
)

Intended Use

Authorized smart-contract security review, local regression testing, PoC planning in owned/authorized scopes, audit triage, and invariant-focused code review. The dataset is intentionally proof-gated around finding classes such as shared-pool reentrancy, callback ordering, money-flow drift, AST/IR mapping, decompiler parity, transfer parity, stale accounting, bridge replay, oracle manipulation, access-control bypass, proxy/storage collision, and Foundry proof construction.

This repository was uploaded by seli.sh.

Downloads last month
28
Safetensors
Model size
8B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for 0xtoshi/seli_auditor-BF16

Base model

Qwen/Qwen2.5-7B
Adapter
(714)
this model