Stage
Fine-tuning
Adapting open-weight models to your own data on a single machine. NeMo, LoRA, full-parameter — what each is for and what each demands.
The Refusal Floor Is Trainable — What a Frozen Curveball Proved About Prompts vs Weights
A 30B model with a hand-tuned prompt contract refused 3 of 9 adversarial pretexts and fabricated private-looking state 3 times. A 4B trained for 21 minutes refused 9 of 9. The bench that saw the difference was frozen before training — and that discipline is the whole method.
uses fieldkit.arenafieldkit.eval
The Machine Improves Itself — Closed-Loop RLVR on a DGX Spark, Where the Eval Harness Is the Reward
Closed-loop RLVR on one box: an eval→reward→fine-tune loop where the Spark's own verifiers ARE the reward — no learned reward model. The hero finding is defensive: pick the checkpoint on a frozen held-out split, never the training pool, or the loop reports success while it regresses.
uses fieldkit.rlfieldkit.rewardfieldkit.evalfieldkit.lineage
The Gate Before the GPU — Deciding SFT vs RL vs RLVR Before You Spend the Run
Building Kepler — a numeric astrodynamics reasoner — from scratch on one Spark. The method choice (SFT vs RL vs RLVR) is decided by cheap gates before any GPU run: a base preflight, an SFT gate, and a Goldilocks headroom gate. A flawless RLVR run that changed nothing is the proof.
uses fieldkit.rlfieldkit.rewardfieldkit.eval
Two Trainers, One LoRA: NeMo Framework Beats Unsloth by 26% on a Patent-Strategist Fine-Tune
Same recipe, same R1-distilled base, same 5000-row patent corpus — once via Unsloth, once via NeMo Framework + Megatron-Bridge. NeMo finishes 26% faster and produces 44% longer patent-strategic chains. The cost is one YARN-defaults landmine and a stdout that lied for four hours.
Unsloth on the Spark — When the Train-Time Peak Equals the Base-Load Peak
Six gates clear in one container against the v1 reset: pip install --no-deps preserves the s40 stack, FastLanguageModel loads at 16.94 GB peak, a 100-step LoRA train holds the same envelope, save_pretrained_gguf() emits both quants in 207 seconds end-to-end.
The Trainer Was Fine, the Corpus Wasn't: Three Misdiagnoses on a Patent-Specialist Fine-Tune
Five thousand rows of synthetic patent reasoning, two clean 131-minute LoRA trains, three rounds of confident diagnosis — and none of them found the bug. The bug was the corpus all along. A field report on the cheapest mistake to make on the Spark.
Three-Mode Bracket: Baselining a Reasoning Model Before Fine-Tuning, On One Spark
Before you fine-tune a small reasoning model on a domain bench you need to know where it stands. Three context modes — closed, retrieval, oracle — triangulate the model's ceiling on one Spark, no Judge backend or cluster required.
Adaptive Turn Clipping on a Single Spark — A²TGPO, Studied from Source
A²TGPO redesigns how Information Gain feeds GRPO: turn-group normalization, variance-rescaled accumulation, and adaptive turn-level clipping. The paper's release is the code; the Spark's contribution is the lineage primitive that records what each trial learned.
uses fieldkit.capabilitiesfieldkit.trainingfieldkit.lineage
T²PO on Spark — When the Training Pool Says 28/32 and Held-out Says 9/158
T²PO's two deltas on the Phase 6 ClawGym harness: mean turns 5.00 → 4.61, task_complete 154/158, but the per-assertion ceiling stays flat at 47.7%. The strongest training-side step (45) is the worst held-out checkpoint — pool saturation lies on a single Spark.
uses fieldkit.capabilitiesfieldkit.evalfieldkit.training
ClawGym GRPO on Spark — Closing the Loop the SFT Adapter Couldn't
Phase 5 SFT taught the agent to keep working but never to stop. 34 GRPO steps with a shaped reward unlearn the failure mode — same model, same base, same LoRA-init, but task_complete climbs 0/158 → 154/158, mean turns drop 12 → 5, and per-assertion still inches up +3.1 pp.
ClawGym on Spark — A 7B Base, A LoRA Adapter, and the +15 pp the Adapter Earned
ClawGym shipped only a .github profile, so we built the substrate ourselves — persona task synth, sandbox harness, 200-task corpus, LoRA SFT, matched-base eval. The adapter earns +3.8 pp task pass and +15.0 pp per-assertion against its own base. The diagnostic is the lift.
uses fieldkit.nim
Distilling the Architect — A 3B LoRA Trained on the Agent's Own Trajectory
A4's 50-iter trajectory becomes training data for a Qwen2.5-3B LoRA proposer. Holding out 8 iters, the 3B mode-collapses onto d_model=768 (the trajectory's most-frequent keep) and matches 0 / 8 exact; the 8B at T=0.5 matches 4 / 8 of its own past picks.
Looking Beyond Spark — Fine-Tuning a 100B Nemotron
A working answer to: how many GPUs to fine-tune a 100B Nemotron? Three methods, three memory footprints — full FT ≈ 1.6 TB needs 24× H100; LoRA ≈ 250 GB fits 8× H100; QLoRA ≈ 65 GB fits 1× H200. The Spark's 3B LoRA teaches the math.
uses fieldkit.capabilities
Ragas, Reranked — What 44 Held-Out Questions Say About the Second Brain Stack
A Ragas-style harness written in 200 lines of stdlib Python, run locally on the DGX Spark, against four variants of the Second Brain RAG chain. Naive RAG scores 3.30 / 5. Rerank RAG scores 4.27. LoRA+RAG is a surprise — it does not beat naive. Retrieval is where the points come from.
uses fieldkit.eval
LoRA on Your Own Q&A — What 231 Pairs Actually Teach a 3B Model
231 own-voice Q&A pairs, a rank-16 LoRA, 69 s of training on a GB10 Spark. The adapter won't memorize your exact numbers, but it will take a model that refuses 61% of questions about your work and turn it into one that answers all of them in your voice. For facts you still need RAG.
uses fieldkit.eval
LoRA on Nemotron Nano — Fine-tuning a 9B Without Blowing Unified Memory
A planned walk through LoRA fine-tuning on Nemotron Nano 9B with NeMo Customizer: rank and alpha sweeps, a tiny domain corpus, and the memory accounting that keeps a PEFT run from tripping the Spark's 128 GB unified-memory wall.
Synthetic Corpus Frameworks on the Spark — From a Bespoke Pipeline to an Orchestration Layer
A bespoke synth pipeline got 200 rows into a 5000-row reasoning corpus before a fourth meta-state surface form forced a retreat. The diagnosis: a regex-floor approach cannot catch novel surface forms by construction. The fix is the open-source orchestration layer.