Describe incremental filtered SFT delivery
Browse files
README.md
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
tags:
|
| 4 |
+
- rl-forgetting
|
| 5 |
+
- filtered-sft
|
| 6 |
+
---
|
| 7 |
+
# Coding SFT baseline delivery, October 27 2026
|
| 8 |
+
|
| 9 |
+
Incremental delivery from the two missing Llama coding SFT baseline arms (the other 13 requested arms were previously delivered to mbpp-code-sft-artifacts). Incomplete runs must not be treated as complete studies. Check the dataset delivery_manifest.json for verified upload units and hashes. Model checkpoints are under runs/checkpoints/step_N/model; final models under runs/final. Load with the corresponding subfolder.
|
| 10 |
+
|
| 11 |
+
Built with Qwen: Qwen2.5-3B derivatives retain the upstream Qwen RESEARCH LICENSE AGREEMENT. Built with Llama: Llama derivatives use Llama 3.2 and subject to the Llama 3.2 Community License and Acceptable Use Policy; see https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct. No additional rights to base models are granted.
|