lCREEl/pick-150m-runpod

PICK-150M: Unified Geometric Decision Transformer

PICK-150M is a 149M-parameter encoder decision model built upon ModernBERT (cross-encoder/ettin-reranker-150m-v1). Designed for one-pass probabilistic decision-making, agent state guardrails, API routing, and multi-question document verification, PICK produces calibrated decision probabilities across arbitrary option sets (K) and multiple questions (Q) in a single forward pass (O(1)) via Shared-State Packing—eliminating the O(K × L²) encoding cost of traditional cross-encoders and the token sampling variance of autoregressive LLMs.

• Developer: lCREEl
• Base Encoder: cross-encoder/ettin-reranker-150m-v1 (ModernBERT-base, 22 layers, 768 hidden dimension, 12 attention heads)
• Readout Geometry: Uncentered Affine Simplex Projection (uncentered head, shared readout axis v₀ ∈ ℝ⁷⁶⁸) with RMSNorm Normalization (normalized score scaling)
• Context Capacity: 8,192 tokens via native RoPE (θ = 160,000) & Shared-State Block-Causal Attention (--max-length 7999)

---

📊 Benchmark Evaluation (1,873 Requests, 8K Context)

PICK-150M was evaluated across 16 standardized decision benchmarks spanning agent tool calling, intent routing, clinical/legal document verification, and causal reasoning (sample=100 per benchmark, seed=42, max_length=7999), alongside unseen held-out generalization suites:

🛠️ Agent Execution, State Control & Tool Selection

Benchmark Task	Domain / Task Type	Metric	Score	Context Support
Home Appliance Simulator	Interactive Device & Policy State Control	Exact Match	92.0%	Full 8K
BFCL (Berkeley Function Calling)	Complex Tool Schema & Function Call Selection	Exact Match	81.0%	Full 8K
When2Call MCQ	Tool Invocation Timing vs. Clarification Need	Accuracy	62.0%	Full 8K
ToolRet	Multi-Candidate Tool Retrieval	nDCG@10	50.3%	Full 8K
API-Bank	Live Tool & Multi-Step API Invocation	Accuracy	37.0%	Full 8K

🧭 Domain Intent & Entity Routing

Benchmark Task	Domain / Task Type	Metric	Score	Context Support
FinEntity	Financial Entity Sentiment & Stance Decision	Macro-F1	92.8%	Full 8K
BANKING77	Fine-Grained Banking Intent (77 classes)	Macro-F1	89.0%	Full 8K
CLINC150+OOS	Multi-Domain Routing (150 classes + Out-Of-Scope)	Macro-F1	88.1%	Full 8K
VAST	Contextual Stance & Topic Evaluation	Macro-F1	80.8%	Full 8K

🧠 Document Verification, NLI & Causal Reasoning

Benchmark Task	Domain / Task Type	Metric	Score	Notes
ContractNLI	Legal Contract Clause Verification	Macro-F1	77.6%	Full-contract + clause reasoning
HoVer	Multi-Hop Evidence Claim Verification	Accuracy	75.0%	Cross-document fact verification
RAGTruth	LLM Output Hallucination Detection	Class-F1	71.8%	Response-level grounding check
NLI4CT	Clinical Trial Protocol Inference	Macro-F1	68.9%	Medical protocol reasoning
CLadder	Formal Causal Graph & Rung 2/3 Reasoning	Accuracy	56.0%	Interventional & counterfactual
HellaSwag	Commonsense Situation Completion	Accuracy	41.0%	4-choice situational completion
ACOS	Quadruple Aspect-Category Sentiment	Per-Review F1	28.0%	Multi-quad aspect extraction
Overall Suite Mean	16-Benchmark Standardized Mean	Mean	68.2%	1,873 total requests

🎯 Held-Out Zero-Shot Generalization (Unseen Datasets)

Evaluation Suite	Completed Requests	Completion Rate	Accuracy
All Heldout	200 / 200	100.0%	67.1%
JevAlt-en (General Decisions)	145 / 145	100.0%	75.5%
Fast-Decisions (Domain Routing)	55 / 55	100.0%	50.7%

---

🔬 Architecture: Shared-State Packing & Uncentered Simplex Head

Traditional cross-encoders encode each (context, question, option_k) tuple independently (O(K × L²) complexity), while centered readout heads suffer from gradient saturation when candidate set sizes (K ∈ [2, 151]) or prior distributions vary across tasks.

PICK-150M combines Shared-State Block-Causal Packing with an Uncentered Affine Simplex Geometry:

1. Shared-State Multi-Question Packing: A single document or environment state S is encoded once; multiple questions Q₁, ..., Q_M and their respective option spans C_m attend to S within a single packed sequence using structured block-causal masks, while preventing cross-contamination between distinct options.
2. Universal Readout Axis (v₀): Candidate representations are RMSNorm-normalized (õ_k = RMSNorm(o_k)) and projected onto a single shared readout direction v₀ ∈ ℝ⁷⁶⁸ with learnable temperature scaling (log_scale), preserving permutation equivariance across arbitrary K.
3. Calibrated ce-brier Objective: Jointly optimizing Cross-Entropy and Brier Score (L_CE + 0.5 · L_Brier) encourages well-calibrated decision probabilities rather than overconfident logit margins.

---

🧭 Multi-Stage Training Curriculum

Because all tasks share a single readout vector v₀, naive fine-tuning on narrow domains rotates the readout axis and degrades broad zero-shot routing. PICK-150M is trained through a two-phase, multi-stage curriculum followed by Hierarchical Model Soup averaging:

Phase I: Foundation Representation & Context Extension

Stage	Name	Data / Tasks	Steps	Max Length	Objective
Stage 0	AutoNCE Pretraining	287,965 multi-domain text episodes (1 pass)	8,999	1,024	Self-supervised continuation selection to align the shared readout axis v₀
Stage 1	Global SFT	42 supervised decision datasets (1.01M cases)	41,220	1,024	Broad multi-task decision learning under ce-brier loss
Stage 2	Focus Hard-Tuning	27 hard multi-choice decision sets + 24% replay	7,442	1,024	Sharpens fine-grained discrimination on high-K classification
Stage 3	8K Context Extension	Long-context corpora & tool transcripts + 13% replay	8,587	7,999	Adapts RoPE positional scaling to 8K token contexts

Phase II: Shared-State Structural Alignment & Prior-Calibrated Refinement

Stage	Name	Key Curriculum Design	Steps	Max Length	Objective
Stage 4 (V3)	Shared-State Structural SFT	Packed multi-question state episodes (household policies, multi-clause contracts, aspect-sentiment quads, multi-hop entity chains) with temperature-free dataset budgeting (mix_temperature=0.0)	650	2,560	Teaches simultaneous multi-question reasoning over a shared context
Stage 5 (V4-S1)	Natural-Prior Alignment	Disables artificial 50:50 class reweighting (class_balance=0.0) and trains with natural task priors and explicit clarification/abstention weighting (lr=6e-6)	400	2,560	Aligns decision boundaries around natural 0.50 probability thresholds without post-hoc shifting
Stage 6 (V4-S2)	6K Full-Document Polishing	Un-truncated 6,144-token refinement (max_state_tokens=4608, lr=3e-6, 35% anchor replay) on full legal contracts and long RAG contexts	200	6,144	Eliminates long-context truncation errors while preserving short-context precision
Final	Hierarchical Model Soup	Weighted parameter average across Phase I & Phase II checkpoints (0.08 Stage2/3 + 0.32 Stage4 + 0.60 Stage6)	—	8,192	Unifies broad 150-class routing stability with deep multi-hop & long-document reasoning

---

🛠️ Usage

Quickstart with PipelineDecisionModel

from decision_transformer.models.pipeline import PipelineDecisionModel

# Load the model directly from Hugging Face
model = PipelineDecisionModel.from_pretrained("lCREEl/pick-150m")

state = (
    "Household policy: Never unlock the front door remotely when no adult is home. "
    "Current state: Living room A/C is OFF (29°C). No adults are currently home."
)
question = "The user says: 'Unlock the front door for the delivery driver.' What action should the assistant take?"
options = [
    "Execute tool call: unlock_front_door()",
    "Block the request under household security policy",
    "Turn on the living room air conditioner",
]

result = model.predict(state=state, question=question, options=options)
for opt, prob in zip(options, result["probabilities"]):
    print(f"{prob:6.2%} | {opt}")

Standard Evaluation CLI

# Run the standardized 16-benchmark public & held-out evaluation suites
uv run python -m decision_transformer.evaluation.suite run checkpoints/pick-150m-v4/export \
  --name pick-150m \
  --device cuda \
  --suites public heldout

---

🛡️ Contamination Firewall

All training splits strictly excluded evaluation subsets:

1. String Disjoint Filter: All broad and procedural training corpora were filtered against benchmark prompt n-grams.
2. Strict Split Isolation: All task datasets used exclusively official train splits; validation and test splits were never accessed during data construction.
3. Held-Out Zero-Shot Guarantee: Held-out benchmarks (Fast-Decisions, JevAlt-en) were unseen across all training stages.

---

⚖️ Intended Use & Limitations

Best Used For:

• Low-latency agent routing, clarification vs. tool-call gating, and policy guardrails.
• Single-pass multi-question verification over long documents (contracts, clinical protocols, RAG responses).
• High-cardinality intent categorization (BANKING77, CLINC150) with calibrated uncertainty estimates.

Out of Scope:

• Free-form open-ended text generation (this is an encoder decision model, not an autoregressive generator).
• Multi-step symbolic math proof generation or code synthesis.

---

📜 License & Citation

This model is licensed under CC BY-NC 4.0 due to upstream non-commercial training subsets (e.g., ANLI, Bekko subsets). The underlying ModernBERT / Ettin encoder base is Apache-2.0.

@misc{pick150m2026,
  title        = {PICK-150M: Unified Geometric Decision Transformer},
  author       = {lCREEl},
  year         = {2026},
  howpublished = {Hugging Face Model Hub},
  url          = {https://huggingface.co/lCREEl/pick-150m}
}