NanoChat.NET — engine layer

The .NET port's engine capabilities, as reachable from code.

12 capabilities 5 sections extracted from Engine capabilities (Dpp.Nano) source: D:\Dev\Lab\NanoChat.NET

Build against it

# give your agent the catalog — no account, no install
claude mcp add provespec -- npx -y provespec-mcp

# then, from any agent session
spec_get(slug: "nanochat-net-engine")                → the capability tree
spec_checklist(slug: "nanochat-net-engine")          → the questions to answer
spec_grade(slug: "nanochat-net-engine", grades: [...])  → parity % + the gap list

# or just fetch it
curl https://provespec.com/catalog/nanochat-net-engine/spec.json

Grade every capability below as yes · partial · no · na, with a file reference as evidence. Whatever you leave ungraded counts as missing — the gap list is the work queue.

Or just paste this to your agent

Use the ProveSpec MCP server. Fetch the "nanochat-net-engine" spec with
spec_get, then walk spec_checklist and grade this repository against it —
one verdict per capability, each with a file reference as the note.
Call spec_grade with the results and give me the gap list as the plan.

If the server is not registered, add it first:
  claude mcp add provespec -- npx -y provespec-mcp

The specification

Data, tokenization, and artifacts 3

✓ Dataset acquisition and streaming Download and integrity-manage ClimbMix shards, stream native Parquet row groups, pack isolated token sequences, partition work by rank, and resume deterministic data cursors.
✓ Tokenizer lifecycle Train the pinned BPE tokenizer from streamed source data, import/export ranks and special tokens, encode/decode batches and boundaries, and preserve exact token bytes and Unicode.
✓ Checkpoint and model-artifact lifecycle Save and resume atomic training checkpoints, warm-start model and optimizer state, publish/load native artifacts, and import/export the upstream Python checkpoint/tokenizer contract.

Training 4

✓ Base pretraining Train NanoGPT with pinned model architecture, Muon/AdamW parameter groups, warmup/warmdown schedules, gradient accumulation, validation, checkpoints, CPU or CUDA execution, and configurable precision.
✓ Supervised task and conversation training Train, resume, and evaluate Generation, Classification, Embedding, Pairwise Ranking, and Conversation/SFT workflows with deterministic task data composition and assistant-only masking.
✓ Reinforcement-learning policy optimization Generate GSM8K rollouts, calculate exact-answer rewards, evaluate pass-at-k, and optimize token-normalized policy gradients through the RL training command.
✓ Distributed and accelerated execution Select float32/BF16/FP16/FP8 CUDA compute, coordinate multi-process ranks and overflow/checkpoint decisions, bind ranks to devices, and select FA3 or the portable attention fallback.

Inference and interaction 2

✓ Inference and generation Load a native model artifact and run cached prefill/decode, reproducible sampling, streaming, cancellation, shared-prefix batch generation, and Generation/Classification/Ranking/Embedding task operations.
✓ Conversation and tool execution Render canonical role/terminal tokens, run multi-turn chat, and service allowlisted calculator or constrained generated-Python tool calls.

Evaluation and performance 2

✓ Quality evaluation Run BPB, CORE, ChatCORE, GSM8K, ARC, MMLU, and HumanEval evaluation with integrity checks, categorical formatting, constrained program execution, per-task metrics, and aggregate scoring.
✓ Inference performance measurement Measure configurable prompt/batch prefill and decode latency, token throughput, model/KV cost, observed memory, and hardware-relative MFU/MBU where a known profile exists.

Official orchestration 1

✓ Official pipelines and experiment sweeps Run resumable official speedrun stages, CPU training, scaling-laws and miniseries matrices, and post-train BPB/CORE sweep evaluation with durable progress/results.

Open in Parity → Download spec.json All 102 specs