Epicenter Glimmer 30B · root-cause investigator (LoRA)

A LoRA adapter for Meta's open-weight Muse Glimmer 30B that turns it into the investigating agent of Epicenter: given a flagged microservice outage, it queries metrics, logs, traces and the call graph through read-only tools, then submits five ranked root-cause services, each citing the tool calls that support it.

It was taught by Meta's hosted Muse Spark agent: the adapter is trained on Spark's investigations of benchmark outages that Spark solved and that Epicenter's deterministic evidence verifier accepted.

Results

Same agent, tools, verifier and 468 held-out test outages from RCAEval (RE1 metrics-only, RE2 and RE3 with logs and traces), served by vLLM from one set of base weights so base and adapter differ only by the adapter. AC@k: the true root-cause service is ranked within the first k.

Glimmer 30B on the test split Base With this adapter
AC@1 · RE1 / RE2 / RE3 0.653 / 0.586 / 0.580 0.911 / 0.944 / 0.815
AC@3 · RE1 / RE2 / RE3 0.751 / 0.605 / 0.642 0.987 / 0.988 / 0.951
Investigations that failed to submit a ranking 140 of 468 1 of 468
Median turns per investigation 11 6
Median model seconds per investigation 126 80

Adapter vs base Glimmer: the gain is significant on every suite (exact McNemar on AC@1, tuned vs base: p = 8e-14, 2e-14 and 3e-4 for RE1, RE2 and RE3). Adapter vs the hosted Muse Spark agent that taught it: the adapter is level on RE2 (0.944 vs 0.963, p = 0.55) and behind on first place for RE1 (0.911 vs 0.951, p = 0.035) and RE3 (0.815 vs 0.975, p < 0.001); its top-3 accuracy is not significantly different from Spark's on any suite.

Every number comes from recorded runs in the Epicenter repository and can be reproduced offline from its model-call cache (scripts/run_explain.py --ranker glimmer-tuned-agent-all --offline).

Training

  • Data. Hosted Muse Spark (via the Meta Model API) investigated every train-split case of RCAEval RE1–RE3. The 246 of 267 runs where Spark ranked the true cause first and the verifier accepted that cause were kept: 212 for training, 34 for validation. Test cases were never read. Each investigation is rebuilt as a chat with Glimmer's own chat template; only the model's turns are trained: its short reasoning notes, its tool calls in Glimmer's native format, and its end-of-turn tokens (403k trained tokens of 2.55M per epoch). Spark's full reasoning is encrypted by its API and was not available to learn from.
  • Method. QLoRA following the recipe in Meta's Muse Glimmer post-training guide: 4-bit NF4 base with double quantization and bf16 compute; LoRA rank 32, alpha 64, dropout 0.05 on the language model's attention (q, k, v, o) and MLP (gate, up, down) projections only, leaving the vision tower untouched; AdamW at 1e-4 with cosine decay and 10% warmup; effective batch 8; 2 epochs (54 steps); sequences up to 32,768 tokens with FlexAttention; loss computed only at trained positions.
  • Compute. One H100 on Modal, peak 66 GB. Validation loss 0.349 → 0.174 → 0.168.
  • Libraries. transformers 5.16.1, peft 0.20.0, bitsandbytes 0.50.2, accelerate 1.13.0, torch 2.12.1.
  • Base. meta-models/Muse-Glimmer-30B at revision a4e59da52a7bc87ae7251dd5545c0dd437c44b68.

The full pipeline is in the repository's scripts/finetune/.

Use

The adapter is meant to run inside Epicenter's agent loop, with its tool schema (top_deviations, compare_to_baseline, get_metric_window, get_logs, get_trace_errors, get_dependencies, submit_ranking). It is not a general chat model.

Serve it with Meta's vLLM image for Glimmer, base and adapter together:

vllm serve meta-models/Muse-Glimmer-30B --revision a4e59da52a7bc87ae7251dd5545c0dd437c44b68 \
  --served-model-name muse-glimmer \
  --enable-auto-tool-choice --tool-call-parser muse_glimmer --reasoning-parser muse_glimmer \
  --generation-config auto \
  --enable-lora --max-lora-rank 64 --lora-modules glimmer-tuned=<path-to-this-adapter> \
  --limit-mm-per-prompt '{"image": 0, "video": 0}'

--limit-mm-per-prompt serves the model text-only; with LoRA enabled, the image's start-up otherwise profiles the vision encoder and fails. The weights are stored in fp32, as trained.

Limitations

  • Benchmark outages are cleaner than production: RCAEval injects one fault at a time into three demo systems. The adapter has not been validated on production incidents.
  • It learned actions, not full reasoning: the teacher's encrypted reasoning wasn't available, and the adapter still trails the hosted model on first place for RE1 and RE3.
  • Tied to Epicenter's tools: its gains come largely from following the investigation protocol (calling tools, citing evidence, submitting on time); with different tools or prompts it may not carry over.
  • Evaluated on a rented GPU. Glimmer is small enough to serve on a single GPU, so a team could run it on hardware it controls; the reported runs were served on Modal.

Data and licensing

Released under Apache 2.0, the base model's license. Training data was derived from the public RCAEval benchmark and from Muse Spark's investigations of it; no private or customer data was used. Built for the Meta Global AI Developer Hackathon (October–November 2026).

Downloads last month
10
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for maxkempler/epicenter-glimmer-30b-rca-lora

Adapter
(48)
this model