Coding-Specialized DSpark Speculator for Muse Glimmer 30B

Motivation

Existing speculators for Muse Glimmer are trained on general chat data (OpenPerfectBlend). None have been trained for long-horizon agentic coding, and the community DSpark model card lists agentic decoding as unbenchmarked. Its weakest published subset is tool-call (3.087 → 3.805 accepted length).

This project fine-tunes a speculator on curated on-policy coding and agentic traces, with the goal of raising acceptance length and throughput on that workload relative to generic speculators. Terminal-Bench is used as the evaluation proxy for a coding customer.

Models

• Target: meta-models/Muse-Glimmer-30B (BF16, 30B)
• Official DFlash drafter (baseline): meta-models/Muse-Glimmer-30B-assistant (5-layer)
• Community DSpark (warm start): DaoCloud/Muse-Glimmer-30B-DSpark
• DFlash2 (baseline): z-lab/Muse-Glimmer-30B-DFlash2

Data

• Replay: DaoCloud/Muse-Glimmer-OPB-100K — 99,984 conversations / 148,900 rows, pre-tokenized, on-policy Glimmer
• New training traces: mini-swe-agent over SWE-Gym, repos disjoint from eval: Satgoy152/Muse-Glimmer-SWE-Gym-2k
  • 2000 instances, balanced round-robin across SWE-Gym's 11 repos.
  • Every instance is verified with a pre-built docker image.
• Eval: Terminal-Bench — Satgoy152/Muse-Glimmer-Terminal-Bench-Eval
  • 40 of 241 tasks, mixed across difficulty (11 easy / 20 medium / 9 hard).
  • Within each difficulty band, tasks are drawn round-robin over category (e.g., games, math, file_operations).
• Leakage control: the training and eval sets are checked for repo overlap before sampling.

Baseline

Official DFlash drafter at 15 speculative tokens, measured on the frozen 40 with
--concurrent 8: pooled acceptance length 3.913, draft acceptance rate
0.194, 151.1 tok/s over 1,753 calls, and 19/39 tasks resolved (48.7%).
Acceptance falls as reasoning strength rises, from 4.345 at low to 3.797 at
xhigh. Per-segment numbers, caveats and coverage:
[benchmark/termina