Datasetpublic

Satgoy152/Muse-Glimmer-SWE-Gym-2k

Muse-Glimmer-SWE-Gym-2k Agentic coding traces from meta-models/Muse-Glimmer-30B, recorded for training a speculative-decoding drafter. 1,981 mini-swe-agent trajectories over SWE-Gym and SWE-bench-extra instances, and the 159,999 individual chat-completion calls behind them. Configs Config Rows Size What it is train 1,981 57 MB One row per trajectory: the full conversation as messages. raw 159,999 2.7 GB One row per recorded API call: request and… See the full description on the dataset page: https://huggingface.co/datasets/Satgoy152/Muse-Glimmer-SWE-Gym-2k.

sourceHugging Faceapache-2.0updated 24d agoView on Hugging Face

2likes257downloads

Dataset Card

Muse-Glimmer-SWE-Gym-2k

Agentic coding traces from meta-models/Muse-Glimmer-30B, recorded for training a speculative-decoding drafter. 1,981 mini-swe-agent trajectories over SWE-Gym and SWE-bench-extra instances, and the 159,999 individual chat-completion calls behind them.

Configs

Config	Rows	Size	What it is
`train`	1,981	57 MB	One row per trajectory: the full conversation as `messages`.
`raw`	159,999	2.7 GB	One row per recorded API call: request and response, for audit and re-rendering.

python

from datasets import load_dataset

train = load_dataset("Satgoy152/Muse-Glimmer-SWE-Gym-2k", "train", split="train")
raw   = load_dataset("Satgoy152/Muse-Glimmer-SWE-Gym-2k", "raw",   split="train")

`train` schema

Column	Type	Notes
`id`	string	Trajectory id — `sha1(first_user_message)[:16]`. Joins to `raw.id`.
`reasoning_strength`	string	`low` / `medium` / `high` / `xhigh`.
`messages`	list<struct>	`role`, `content`, `reasoning_content`, `tool_call_id`, `tool_calls`.
`tools`	string	JSON. Constant across every row: a single `bash` function.

The last message in messages is the assistant's final reply, so the row is a complete conversation. content is null on assistant turns that only emit tool calls.

Generation


Target model	`meta-models/Muse-Glimmer-30B`
Serving	vLLM 0.28.1rc1
Drafter	dflash, `meta-models/Muse-Glimmer-30B-assistant`
`num_speculative_tokens`	15
Prefix caching	on
Sampling	temperature 1.0, topp 0.95, topk 64
Agent	`mini-swe-agent`, single `bash` tool

Effort split

Level	Trajectories	Calls
low	593	42,529
medium	596	48,648
high	594	51,606
xhigh	198	17,216

Speculator baseline

Metric	Median	Mean	Range
`acceptance_length`	4.974	5.046	3.862 – 9.917
`draft_acceptance_rate`	0.265	0.270	0.191 – 0.594
`prefix_cache_hit_rate`	0.978	0.975	0.480 – 0.994

Pooled over the run: 3,814,699 accepted of 14,579,910 drafted across 971,994 draft steps — acceptance rate 0.2616, acceptance length 4.925.

License

apache-2.0. Derived from SWE-Gym and SWE-bench-extra; upstream instance licenses apply.