Update dataset card with clean/unfiltered subsets

README.md

@@ -18,15 +18,40 @@ tags:
 - dclm
 pretty_name: WebTerminal
 size_categories:
-- 10M<n<100M
+- 1M<n<10M
+configs:
+  - config_name: clean
+    data_files:
+      - split: train
+        path: clean/*.parquet
+    default: true
+  - config_name: unfiltered
+    data_files:
+      - split: train
+        path: unfiltered/*.parquet
 ---
 # Terminal/CLI Web Text
 
 ![webterminal](webterminal.png)
 
-**v0.1**: 360B tokens, really noisy
+A filtered extract of terminal and command-line content from two large web-text corpora, designed for upsampling agentic-adjacent data during pretraining.
 
-A filtered extract of terminal and command-line content from two large web-text corpora, consists of 360B tokens
+## 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 |
+
+```python
+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
 
@@ -35,6 +60,8 @@ A filtered extract of terminal and command-line content from two large web-text
 
 ## 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:
 
@@ -48,28 +75,67 @@ A filtered extract of terminal and command-line content from two large web-text
 
 Documents scoring >=5 are kept.
 
-3. **Dedup**: exact dedup across both datasets using xxhash64 on full text.
+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 `$` matches currency amounts, and indented Python/SQL code gets scored as terminal content.
+
+v0.2 applies a three-stage structural filter over the unfiltered data:
+
+1. **Context-aware gate**: instead of matching bare `$`, 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% genuine 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
 
-| | Chunks | Size | Rows |
-|---|---|---|---|
-| DCLM | 13,144 | ~229 GB | ~18.8M |
-| FineWeb | 8,800 | ~669 GB | ~47.5M |
+### Clean (v0.2)
 
-| Score | Count | % | Cumulative |
-|---|---|---|---|
-| 5 | 39,025,201 | 63.62% | 63.62% |
-| 6 | 10,787,199 | 17.59% | 81.21% |
-| 7 | 4,063,886 | 6.63% | 87.83% |
-| 8 | 2,911,983 | 4.75% | 92.58% |
-| 9 | 1,304,162 | 2.13% | 94.70% |
-| 10 | 1,022,996 | 1.67% | 96.37% |
-| 11-14 | 1,609,090 | 2.62% | 98.99% |
-| 15-20 | 536,421 | 0.87% | 99.87% |
-| 21-34 | 80,340 | 0.13% | 100.00% |
-| **Total** | **61,341,278** | | |
+- **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
 
-Mostly for upsampling agentic-adjacent data during pretraining.
+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.