MetisRCA-E4B: A 4B-Class Root-Cause Model That Cites Evidence or Steps Back
This is for teams attaching root-cause analysis to agents. MetisRCA-E4B reaches 0.873 top-1 on a held-out set and cuts unsupported diagnoses to 0.3%. On failure mechanisms it has never seen it trails a same-size Qwen3.5-4B, and it abstains unnecessarily on about 15% of determinable cases.
What Nemotron 3.5 Bolt E4B is
Bolt E4B is a small language model NVIDIA post-trained from Gemma-4-E4B-it. It has 8B total and 4B effective parameters and up to 64K tokens of input and output. NVIDIA strengthens multi-turn tool use, instruction following and input grounding, and trains with GRPO reinforcement learning. Low memory and low time-to-first-token are the design goals, which suits local assistants and edge devices. The license is OpenMDW-1.1, and a separate 4GB Q4_K_M build is published. The BF16 release is meant for customization, so we used it as our base.
How it was built
Figure 1. The two yellow stages keep scores honest. Label generation and leakage checks belong to code, not to a model.
We ran LoRA SFT on the BF16 weights: rank 16, alpha 32, adapters on 258 linear layers across the attention and MLP text path, learning rate 2e-4, 2 epochs. Gold labels come from code, and we ran leakage and lineage checks between train and evaluation sets. The model reads incident evidence and names a root cause. When evidence is insufficient it answers INSUFFICIENT_EVIDENCE and asks for the next tool, in a read-only role.
Measured results
We compared three models with identical prompts and decoding. Each row pairs values measured on the same evaluation rows.
| Evaluation | MetisRCA-E4B | Qwen3.5-4B | Bolt base |
|---|---|---|---|
| Held-out set overall (1,566 cases), root cause top-1 | 0.873 | 0.694 | 0.619 |
| Never-seen mechanisms (783 rows), top-1 | 0.784 | 0.814 | - |
| Cloud-OpsBench, determinable never-seen mechanisms, top-1 | 0.783 | 0.831 | 0.764 |
On the overall held-out set, evidence F1 is 0.82 and unsupported diagnoses are 0.3%, against 7.3% for Qwen3.5-4B. The gap is larger in behavior than in accuracy. Stepping back when evidence is thin is what this adapter is for.
Calibration: an abstain policy
The repo ships an inference-time abstain policy, abstain_policy.py, with the same weights. The threshold of 1.5 was chosen on validation only and applied once to the test set.
| Never-seen mechanisms (783 rows) | Greedy | Calibrated |
|---|---|---|
| Root cause top-1 (Qwen3.5-4B 0.814) | 0.784 | 0.824 |
| Over-abstain | 15.3% | 9.4% |
| Correct abstain | 78.7% | 73.6% |
| Unsupported diagnoses | 0.26% | 0.26% |
The pre-registered gate of over-abstain below 5% was not met. On the 360 determinable Cloud-OpsBench rows, compared like for like, greedy gets 282 right, calibrated 298 and Qwen3.5-4B 299, one row ahead. Top-1 on the whole held-out set rises from 0.873 to 0.897. Under llama.cpp the GGUF behaves like greedy v4, because calibration needs transformers. The Q4_K_M GGUF (5.34 GB) keeps top-1 0.9697 on the seen split, the same as BF16.
Honest limits
On never-seen failure mechanisms Qwen3.5-4B scored higher under greedy decoding, and we do not claim the calibrated model finds new causes better. Its strength is evidence grounding and abstention. About 15% of determinable cases end in an abstention. The adapter tends to stop rather than invent a cause, which is the safer failure for an agent, but a new class of incident needs a path to a larger model or a person.
How to use it in Paxis
Paxis is ThakiCloud’s platform for automating enterprise work with AI agents. What follows is a design direction, not a shipped feature. MetisRCA-E4B sits as a read-only RCA step. It cites evidence ids for a cause, or abstains and requests the next tool it needs.
Figure 2. The model proposes and code holds permission. The RCA step is read-only, so an abstention leads to a request for the next tool.
If an abstention leads to one more lookup by the agent, the roughly 15% over-abstention costs one extra tool call. Unfamiliar incident types go to a larger model or a person. Which remediation runs is decided by a permission check outside the model, not by this model.
Get the model
- MetisRCA-E4B: https://huggingface.co/ThakiCloud/MetisRCA-E4B
- Collection: https://huggingface.co/collections/ThakiCloud/thakicloud-bolt-e4b-specialists-6ade97db98b0258cb49447a0
