docs: add a card stating what is actually in this repo
Browse files
README.md
ADDED
|
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags:
|
| 4 |
+
- polly
|
| 5 |
+
- scbe
|
| 6 |
+
- qwen2
|
| 7 |
+
- text-generation
|
| 8 |
+
- sft
|
| 9 |
+
- merged
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
<!-- scbe-status -->
|
| 13 |
+
> **Status: superseded.** Research artifact, not a production candidate. Canonical model: [scbe-coding-agent-vtc-qwen15-v1-gguf](https://hf.co/issdandavis/scbe-coding-agent-vtc-qwen15-v1-gguf).
|
| 14 |
+
|
| 15 |
+
# Polly r8 - Qwen2.5-Coder-1.5B (merged)
|
| 16 |
+
|
| 17 |
+
Full-weight merge. Facts below are from `MERGE_INFO.txt` in this repo.
|
| 18 |
+
|
| 19 |
+
- **Base**: `Qwen/Qwen2.5-Coder-1.5B-Instruct`
|
| 20 |
+
- **Adapter**: `training/runs/local_1.5b_r8_20260415T202900/checkpoint-500`
|
| 21 |
+
- **Round**: r8 - 137K coding examples + tokenizer + spiral seal
|
| 22 |
+
- **Method**: `merge_and_unload` (LoRA baked into base weights)
|
| 23 |
+
|
| 24 |
+
## Important
|
| 25 |
+
|
| 26 |
+
The Kaggle 0.5B adapter at
|
| 27 |
+
[polly-r8-qwen-0.5b](https://hf.co/issdandavis/polly-r8-qwen-0.5b) is a
|
| 28 |
+
**different base model**. The two cannot be averaged or merged together.
|
| 29 |
+
|
| 30 |
+
## Not published here
|
| 31 |
+
|
| 32 |
+
No eval numbers are included in this repo, so none are claimed. Treat this as
|
| 33 |
+
a training artifact from round r8, not a benchmarked release.
|