Nemotron Pi Agent V5 training report

Plain-English overview

The starting point was NVIDIA's Nemotron Cascade 8B model. Instead of retraining all eight billion parameters, this run used QLoRA: the base model was loaded in a memory-efficient 4-bit form and a small set of trainable adapter weights was attached to it. The adapter is only about 178 MiB, while the underlying model remains unchanged.

The model saw examples of complete coding-agent conversations rendered with Pi 0.82.0's exact system prompt and tool schemas. That matters because an agent must learn two things at once: how to write good code and how to express a tool call in the exact machine-readable shape the host expects.

Base model

nvidia/Nemotron-Cascade-8B at revision 276a779fe0c1fe2d63a68fd573fec7b4b0a1907e

Method

4-bit NF4 QLoRA, rank 32, alpha 64, dropout 0.05

Context limit during training

20,480 tokens

Serving context

32,768 tokens

Reasoning mode

Off; the model is served with Hermes tool parsing only

The V5 dataset

V5 contains 1,118 accepted trajectories and 4,555 assistant actions. Exporting those conversations produced 4,119 training records and 436 validation records. The validation set is kept out of gradient updates so it can measure whether the model learned patterns that generalize beyond the examples used to change its weights.

A full tokenizer audit rendered every record before training: 18,953,799 prompt tokens and 3,438,933 supervised assistant tokens. No prompt token was accidentally included in the loss, no hidden <think> content appeared in the target labels, and the longest sequence was 19,082 tokens—safely below the 20,480-token training cap.

How the training run went

The run completed all 515 optimizer steps in about 5 hours 25 minutes. Evaluation ran before training and every 50 steps. Loss fell sharply by step 50, then continued improving slowly until step 500. The final step was only 0.0000168 worse than the best checkpoint—far too small to claim a meaningful difference—so the simpler final adapter is the served model while checkpoint 500 remains archived.

System health

A separate guard process sampled the machine every 30 seconds for 656 samples. It was configured to stop training after two readings at or above 84 °C, if available host memory fell below 4 GiB, or if free disk fell below 20 GiB. None of those guardrails fired.

Some unusually long sequences took one to five minutes because they forced more memory movement across PCIe. They recovered normally; there were no out-of-memory errors, thermal slowdowns, hardware error flags, or crashes.

Before and after

What a successful V5 run looked like

1. Called read first with a valid object argument.
2. Called edit with an exact replacement that preserved behavior.
3. Called bash with the exact requested svelte-utils checks.
4. Inspected the clean result and then summarized the verified change.

The main failure mode was not ignorance of the workflow. The model sometimes produced the right JSON text but wrapped it as a string, causing Pi's schema validator to reject it. All three create-from-scratch trials produced the requested source and ran validation; two ended cleanly, but both failed the benchmark's stricter requirement that the very first tool be write because the model checked for the missing file first. The remaining trial made the correct file but did not converge to a clean final handoff.

What the training terms mean

Token

A small unit of text. Code punctuation, identifiers, and words are split into tokens the model predicts one at a time.

Loss

A penalty for assigning too little probability to the correct next token. Lower is better when comparing the same model on the same held-out dataset.

Token accuracy

The fraction of supervised positions where the most likely next token exactly matched the target. A single wrong token can still break a JSON tool call, so 98% token accuracy does not mean 98% successful agent tasks.

Epoch

One pass through the training dataset. This run used one epoch to reduce the risk of memorizing the examples.

Batch and gradient accumulation

The GPU processed one example at a time and accumulated gradients across eight examples before each optimizer update, giving an effective batch size of eight.

Validation set

Examples not used to update weights. They detect whether the model improves beyond the exact training records, but only on the same kind of next-token imitation task.

QLoRA adapter

A compact set of learned corrections layered over a quantized base model. It makes fine-tuning feasible on one 32 GiB GPU and leaves the original model intact.

Trying V5 in Pi

The model is served on Morningstar through an OpenAI-compatible endpoint. Pi should use the provider ID local-svelte-agent, model ID pi-svelte-agent-v5, and reasoning disabled.

pi --model local-svelte-agent/pi-svelte-agent-v5

The provider base URL is https://morningstar.otter-hawksbill.ts.net:10001/v1. The checked-in example configuration is training/pi-model-v5.example.json.

The serving wrapper pins enable_thinking=false as a server-side chat-template default. This is necessary because Pi's generic OpenAI-compatible requests do not carry Nemotron's model-specific template flag; without the server default, literal thinking tags can leak into ordinary output.

Recommendation

Keep V5 as the best Nemotron Pi experiment, but do not promote it as a reliable daily driver. The next dataset should target the observed boundary directly: many short examples contrasting object-valued tool arguments with accidentally stringified JSON, more create-file trajectories with efficient first-tool selection and prompt termination, and negative examples that stop verbose internal planning from entering normal output.

After that cleanup, repeat this exact nine-case Pi benchmark before moving the recipe to the larger Qwen model. The larger model should improve capacity, but only if the dataset and serving stack agree on the same tool-call representation.