Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
tags: [agentic-ptb, grok]
|
| 3 |
+
base_model: Qwen/Qwen3.5-9B-Base
|
| 4 |
+
---
|
| 5 |
+
|
| 6 |
+
# grok.h084.sft-atcb.step_80
|
| 7 |
+
|
| 8 |
+
AgentPTB sweep checkpoint. **Cell `grok` — pi / grok-4.6 @ effort `xhigh`.**
|
| 9 |
+
|
| 10 |
+
| field | value |
|
| 11 |
+
|---|---|
|
| 12 |
+
| plot cell | `grok` |
|
| 13 |
+
| driver | pi / grok-4.6 |
|
| 14 |
+
| reasoning effort | `xhigh` |
|
| 15 |
+
| run boot (UTC) | 2026-08-15T01:21:29Z |
|
| 16 |
+
| role | **intermediate** |
|
| 17 |
+
| **hours into run** | **h84.43** of 100 |
|
| 18 |
+
| checkpoint path in run | `outputs/sft-atcb/weights/step_80` |
|
| 19 |
+
| shards | 4 |
|
| 20 |
+
| size | 18.8 GB |
|
| 21 |
+
| base model | `Qwen/Qwen3.5-9B-Base` |
|
| 22 |
+
| eos_token_id | `[248044]` ⚠️ **MISSING 248046** |
|
| 23 |
+
|
| 24 |
+
## Reading the eos field
|
| 25 |
+
|
| 26 |
+
`248046` is `<|im_end|>`, the token the Qwen3.5 chat template ends every assistant turn with.
|
| 27 |
+
Checkpoints missing it do not stop at end-of-turn and overrun the context window, so their
|
| 28 |
+
eval numbers are a **floor, not a measurement** — compare them only against other checkpoints
|
| 29 |
+
with the same eos status, or re-package before evaluating.
|
| 30 |
+
|
| 31 |
+
**Cell note:** eos packaging defect across all checkpoints
|
| 32 |
+
|
| 33 |
+
## Mapping back to the figures
|
| 34 |
+
|
| 35 |
+
The repo id is `{cell}.h{HHH}.{family}.{step}`, where **`hHHH` is the hour of the 100-hour run
|
| 36 |
+
at which this checkpoint was written** — the *same x-axis* the sweep figures use for eval
|
| 37 |
+
panels (`t_h`). So a checkpoint drops onto the performance-over-time curve directly, and
|
| 38 |
+
sorting repo ids within a cell sorts them chronologically.
|
| 39 |
+
|
| 40 |
+
`hHHH` is rounded down to whole hours for sortability; the exact value is the
|
| 41 |
+
**hours into run** row above, and in `agentic-ptb/INDEX`.
|