Datasets:
Varek Phase 0 dataset build
Browse files- README.md +33 -0
- train-00000-of-00004.parquet +3 -0
- train-00001-of-00004.parquet +3 -0
- train-00002-of-00004.parquet +3 -0
- train-00003-of-00004.parquet +3 -0
README.md
ADDED
|
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Varek SFT Mix A - Generalized Coding
|
| 2 |
+
|
| 3 |
+
Part of the **Varek** project: a sub-1B parameter edge-deployable model for generalized coding.
|
| 4 |
+
|
| 5 |
+
## Composition
|
| 6 |
+
|
| 7 |
+
| Source dataset | Conversations |
|
| 8 |
+
|---|---|
|
| 9 |
+
| `qwen3.8-max-50k` | 44,179 |
|
| 10 |
+
| `qgk-sft_code` | 13,427 |
|
| 11 |
+
| `qgk-sft_reasoning` | 9,950 |
|
| 12 |
+
| `multi-distilled` | 8,242 |
|
| 13 |
+
| `qgk-sft_math` | 4,891 |
|
| 14 |
+
| `multi-cybersecurity` | 4,849 |
|
| 15 |
+
| `multi-coding` | 2,229 |
|
| 16 |
+
| `kimik3` | 2,033 |
|
| 17 |
+
| `claude-mythos` | 75 |
|
| 18 |
+
|
| 19 |
+
**Total: 89,875 conversations**
|
| 20 |
+
|
| 21 |
+
## Processing pipeline
|
| 22 |
+
|
| 23 |
+
1. Parsed upstream parquet/jsonl (messages stored as JSON strings were decoded)
|
| 24 |
+
2. OpenAI-format tool_calls normalized (arguments JSON-string -> dict) for Qwen chat-template compatibility
|
| 25 |
+
3. Quality filters: quality score >= 0.75 (fable-5), >= 0.7 (q38/vibecoding), sampling_weight > 0 (qgk)
|
| 26 |
+
4. Cross-source deduplication: exact-conversation SHA-256 + assistant-output-only SHA-256
|
| 27 |
+
5. Test splits of every upstream family held out globally (leak blocklist) - reserved for evaluation
|
| 28 |
+
6. RaceBench excluded entirely per project policy; never loaded or referenced
|
| 29 |
+
|
| 30 |
+
## License
|
| 31 |
+
|
| 32 |
+
Inherits upstream licenses: MIT for most sources. The `multi-*` shards derive from an
|
| 33 |
+
upstream repository published under a custom/other license; treat this derivative accordingly.
|
train-00000-of-00004.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:5f4eabd3841451c6ed6a6c3209bb61fd0707f31ec9e7a831776e7d487c0ca649
|
| 3 |
+
size 36415798
|
train-00001-of-00004.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:32ce2e69d5dae34efc321accdec7094ac698ef5ac7f48e4041a9f2627974293c
|
| 3 |
+
size 26407828
|
train-00002-of-00004.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:20748ac47d16fca5a898052124c86aac5122006c345543e5a170b3567c28e943
|
| 3 |
+
size 28454825
|
train-00003-of-00004.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d2683d4a70d4c2c3476ebc0b65842417519b297b6bcdf4391a70249baf0b8951
|
| 3 |
+
size 21777382
|