Concept Registry — data-science-mcp¶
Prefixes:
CONCEPT:DSCI-*(this project) ·CONCEPT:ML-*(cross-repo LLM trainer) Version: 0.9.0 Bridge:CONCEPT:AU-ECO.messaging.native-backend-abstraction(Unified Toolkit Ingestion)
Project-Specific Concepts¶
| Concept ID | Name | Description |
|---|---|---|
CONCEPT:DS-ECO.mcp.data-management |
Data Management Operations | MCP tool domain data_management — Action-routed dynamic tool registration |
CONCEPT:DS-ECO.mcp.interpretability |
Interpretability Operations | MCP tool domain interpretability — Action-routed dynamic tool registration |
CONCEPT:DS-ECO.mcp.model-evolution |
Model Evolution Operations | MCP tool domain model_evolution — Action-routed dynamic tool registration |
CONCEPT:DS-ECO.mcp.model-training |
Model Training Operations | MCP tool domain model_training — Action-routed dynamic tool registration; incl. the in-house training substrate (training_data corpus/reward engine + trainers/ SFT/DPO/GRPO + peft_manager/tokenizer_registry/rollout_buffer, CONCEPT:AU-AHE.evaluation.adaptive-reasoning-effort) |
CONCEPT:DS-ECO.mcp.quant-statespace |
State-Space / Stat-Arb Operations | MCP tool domain quant_statespace — Kalman filter/beta/volatility, ADF, OU calibration + thresholds, Markov transition (engine client.finance.*, EG-KG.domains.state-space-statistical-arbitrage) |
CONCEPT:DS-ECO.mcp.signal-combination |
Signal-Combination Operations | MCP tool domain quant_signals — order-book imbalance, information ratio, effective independent N, alpha combination, convergence gate; plus empirical_kelly (quant_sizing) and brier_score (quant_validation) (engine client.finance.*, EG-KG.domains.quant-finance) |
CONCEPT:DS-ECO.mcp.sabr-volatility |
SABR Volatility-Surface Operations | MCP tool domain quant_derivatives — Hagan-2002 SABR implied_vol / smile / calibrate (fit α,ρ,ν with β fixed → {alpha,beta,rho,nu,rmse,converged}) delegating to engine client.finance.sabr_* (AU-KG.domains.derivatives) |
LLM Trainer Concepts (CONCEPT:ML-*)¶
The high-caliber LLM trainer — create, pretrain from random init, and fine-tune
models, driven by AI agents. This is a deliberate cross-repo family (it spans
data-science-mcp + agent-utilities + universal-skills), so it uses a repo-neutral
ML-* prefix rather than DSCI-*. It expands CONCEPT:DS-ECO.mcp.model-training
(Model Training Operations) and bridges CONCEPT:AU-AHE.evaluation.adaptive-reasoning-effort
(the in-house training substrate). See Model Training, Installation,
and the SDD spec at .specify/specs/llm-model-trainer/.
| Concept ID | Name | Description |
|---|---|---|
CONCEPT:AU-AHE.trainer.high-caliber-llm-trainer |
Trainer Hardening | Shared trainers/loop.py::run_loop — precision (fp16 scaler / bf16 autocast), gradient accumulation, clipping, LR scheduling, checkpoint save+resume, metrics. SFT/DPO/GRPO/pretrain all route through it; TrainConfig defaults reproduce prior behaviour exactly |
CONCEPT:DS-AHE.trainer.data-engine |
Corpus Curation Engine | data_engine.py (+ mcp/mcp_data_engine.py) — stream / exact+near dedup / decontaminate / quality-filter / pack / DatasetVersion lineage; epistemic-graph HNSW/LSH find_similar_pairs accelerates the all-pairs search (local-cosine fallback) |
CONCEPT:DS-AHE.trainer.concept-2 |
Pretrain From Random Init | tokenizer_trainer.py (BPE) + trainers/pretrain_trainer.py (PretrainSpec, AutoConfig→from_config, packed next-token CE; kind=pretrain) + run_pretrain_pipeline |
CONCEPT:DS-AHE.trainer.concept-3 |
Experiment Tracking | tracking.py::RunTracker — MLflow / W&B / none + best-effort epistemic-graph TrainingRun provenance mirror |
CONCEPT:DS-AHE.trainer.concept-4 |
Distributed Scale-Out | trainers/accelerate_launch.py + launch/ — FSDP and DeepSpeed ZeRO-3 as first-class peers + accelerate launch config/command builder (homelab or cloud) |
CONCEPT:DS-AHE.trainer.concept-5 |
Benchmark Evaluation | trainers/eval_hooks.evaluate_benchmarks — EleutherAI lm-eval scoring alongside the AHE-3.1 reliability suite |
CONCEPT:AU-AHE.trainer.concept-2 |
Agent-Driven Training | Personas data_curator/training_engineer/eval_judge/ml_orchestrator (agent-utilities) + the ml/train_model workflow + model_training_team (universal-skills) |
CONCEPT:DS-AHE.reward.one-sequence-level-score |
Reward Model | trainers/reward_trainer.py (Bradley-Terry pairwise loss objectives.bradley_terry_loss) on a scalar head (trainers/value_head.py); consumes the preference corpus; tool train_reward. The RLHF stage between SFT and PPO |
CONCEPT:DS-AHE.trainer.per-token-value |
PPO (actor-critic) | trainers/ppo_trainer.py — rollout → reward (verifier or DS-AHE.reward.one-sequence-level-score reward model) → GAE (objectives.gae) + value head → clipped surrogate (grpo_surrogate) + value loss (objectives.value_function_loss) + KL-to-reference; tool train_ppo; training_pipeline.run_rlhf_pipeline chains SFT→reward→PPO |
CONCEPT:DS-AHE.trainer.data-transformation |
Flat-Token Pretrain Data | data_engine.prepare_pretrain_data / read_token_blocks — stream (.jsonl/.jsonl.zst/HF) → tokenize (EOS-sep) → contiguous HDF5/.npy token array, batched on the fly (no padding); tool prepare_pretrain_data |
CONCEPT:AU-AHE.trainer.join-inference |
Training-Job Dispatch | training_job_runner.py runs a trainer as a JobRunner on the agent-utilities GPU-slot scheduler (KG-2.65): checkpoint on should_pause, auto-resume on backfill; the train_model workflow submits to a GPU host or falls back to accelerate launch |
CONCEPT:DS-AHE.trainer.chat-format |
Chat + Reasoning Format | chat_template.py — learnable role markers + <think>/<answer> (ordinary tokens) + answer extractor; eval_hooks.evaluate_gsm8k / gsm8k_reward verifiable exact-match reward |
CONCEPT:DS-AHE.optimization.cache-agent-loop |
CacheRL Cached Rollouts + Hybrid Thinking | cache_agent_loop.py — three-tier fuzzy tool-call cache (ThreeTierToolCache exact→fuzzy→semantic) wrapped by CacheAgentLoop so multi-turn RL rollouts serve repeated tool calls from cache (live-execution cost ↓, calls_saved/hit_rate measure it) + ThinkingTraceAugmenter interleaving model "why-this-tool" rationale and per-segment provenance labels; reward half (token-mask + cache-tier-aware reward) in agent-utilities AHE-3.49. Distils arXiv:2606.14179 |
Cross-Project References (from agent-utilities)¶
| Concept ID | Name | Origin |
|---|---|---|
CONCEPT:AU-ECO.messaging.native-backend-abstraction |
Unified Toolkit Ingestion | agent-utilities |
CONCEPT:AU-ORCH.adapter.hot-cache-invalidation |
Confidence-Gated Router | agent-utilities |
CONCEPT:AU-OS.config.secrets-authentication |
Prompt Injection Defense | agent-utilities |
CONCEPT:AU-OS.state.cognitive-scheduler-preemption |
Cognitive Scheduler | agent-utilities |
CONCEPT:AU-OS.governance.reactive-multi-axis-budget |
Guardrail Engine | agent-utilities |
CONCEPT:AU-OS.governance.wasm-micro-agent-sandbox |
Audit Logging | agent-utilities |
CONCEPT:AU-KG.query.object-graph-mapper |
Knowledge Graph Core | agent-utilities |
CONCEPT:AU-AHE.evaluation.adaptive-reasoning-effort |
Training Substrate (reward decomposition / distillation) | agent-utilities |
Synergy with agent-utilities¶
This project integrates with agent-utilities via CONCEPT:AU-ECO.messaging.native-backend-abstraction (Unified Toolkit Ingestion). The data_science_mcp MCP server registers its tools with the agent-utilities FastMCP middleware, enabling automatic discovery, telemetry, and Knowledge Graph ingestion of all DSCI-* concepts.