AdaMLLab/WebTerminal

Terminal/CLI Web Text

A filtered extract of terminal and command-line content from two large web-text corpora, designed for upsampling agentic-adjacent data during pretraining.

Subsets

Subset	Rows	Tokens	Size	Quality
clean (default)	2.33M	4.6B	11 GB	~98% terminal content
unfiltered	61.3M	359B	962 GB	~15% terminal content

from datasets import load_dataset

# Load the clean subset (default)
ds = load_dataset("AdaMLLab/WebTerminal")

# Load the unfiltered subset
ds = load_dataset("AdaMLLab/WebTerminal", "unfiltered")

Sources

• DCLM (Zyphra/dclm-dedup)
• FineWeb (Salesforce/fineweb_deduplicated)

How it was built

v0.1 Unfiltered

1. Fast filter: skip any document that doesn't contain obvious CLI indicators ($, sudo, pip install, ```bash, root@, etc.)
2. Score: remaining docs are scored (0-34) across five signals, each with a per-match point value and a cap:

Filter	Description	Points	Cap
Prompt patterns	Shell prompts like $ cmd, user@host:~$, >>>, root@, PS C:\	2 per match	10
CLI commands	Known commands: sudo, apt-get, pip install, git clone, docker run, curl, ssh, gcc, etc. (30+ patterns)	1 per unique match	8
stdout patterns	Output indicators: "successfully installed", "cloning into", drwx (ls output), "packets transmitted", "traceback", version strings	2 per match	6
Code blocks	Terminal-flavored code blocks: ```bash, ```shell, <pre><code>, terminal/console div classes	2 per match	6
Indented blocks	3+ consecutive lines indented 4+ spaces (code/output blocks)	1 per match	4

Documents scoring >=5 are kept.

3. Dedup: exact dedup across both datasets using xxhash64 on full text. Removed 1,168 duplicates.

v0.2 Clean

The unfiltered subset is ~84-86% noise at lower score levels (5-12), which make up 93% of the data. The root cause: v0.1's scoring uses context-blind keyword matching, CLI command names like find, make, cat appear in normal English prose, bare $ matches currency amounts, and indented Python/SQL code gets scored as terminal content.

v0.2 applies a three-stage structural filter over the unfiltered data:

1. Context-aware: instead of matching bare $, requires $ sudo, $ git, $ docker, etc. (dollar sign + space + known command). Eliminates ~87% of documents immediately.
2. Validation regex: confirms a genuine structural terminal pattern exists, shell prompts followed by real commands, user@host:~$ patterns, Python REPL >>>, tracebacks, ```bash code blocks, Unix file permission listings, man page headers, shebangs.
3. Weighted structural scoring (term_score_v2): each pattern has a weight (1-3) and occurrences are capped. Documents need term_score_v2 >= 3 to be kept.

Weight	Signal	Max
3	Command prompts ($ cmd at line start)	9
3	SSH prompts (user@host:~$)	9
2	Python REPL, file listings, tracebacks, terminal code blocks, git/docker ops, Windows prompts, man pages	2-6 each
1	Install output, systemd units, shebangs, sudo commands	1 each

No indentation-based scoring. No context-blind command substring matching.

Result: 3.8% of the unfiltered data survives, from 61.3M rows down to 2.33M rows. Quality jumps from ~15% to ~98% terminal/CLI content.

Schema

Clean subset

Column	Type	Description
text	string	Document text
term_score	int32	Original v0.1 score (5-34)
term_score_v2	int32	Structural score from v0.2 filter (3+)

Unfiltered subset

Column	Type	Description
text	string	Document text
term_score	int32	Original v0.1 score (5-34)

Stats

Clean (v0.2)

• 2,334,414 rows | 4.6B tokens (Llama-3.2-1B tokenizer) | 11 GB
• 62 parquet files, ~169-185 MB each, snappy compressed

Unfiltered (v0.1)

• 61,341,278 rows | 359B tokens | 962 GB
• 4,187 parquet files, ~180-240 MB each, snappy compressed

v0.1 Score	Count	%
5	39,025,201	63.62%
6	10,787,199	17.59%
7	4,063,886	6.63%
8	2,911,983	4.75%
9-14	3,594,547	5.86%
15-34	958,462	1.56%

Use case

Upsampling agentic-adjacent data during pretraining. The clean subset is recommended for most use cases. The unfiltered subset is available for researchers who want to apply their own filtering.