File size: 5,590 Bytes
420f589 c9e9339 420f589 28c1295 c9e9339 420f589 c9e9339 420f589 c9e9339 420f589 9e4828a 420f589 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 | ---
tags: [agentic-ptb, index]
---
# AgentPTB checkpoint index
Every checkpoint produced by the AgentPTB driver × reasoning-effort sweep, one HF repo each.
All are `Qwen/Qwen3.5-9B-Base` derivatives in standard safetensors format.
## Model id format
```
agentic-ptb/{cell}.h{HHH}.{family}.{step}
```
**`hHHH` is the hour of that cell's 100-hour run at which the checkpoint was written** — the
same x-axis the sweep figures use for eval panels. A checkpoint therefore drops straight onto
the performance-over-time curve, and sorting ids within a cell sorts them chronologically.
`hNA` means the hour could not be recovered (see `hour_source`).
`{cell}` is the plot key, so results join back to the figures directly.
## Cells
| cell | driver | effort | checkpoints |
|---|---|---|---|
| `sol-high` | Codex / gpt-5.6-sol | high | 57 |
| `grok` | pi / grok-4.6 | xhigh | 59 |
| `dpsk-v4-flash` | pi / DeepSeek v4-flash | thinking | 40 |
| `kimi` | kimi-code / kimi-k3 | high | 38 |
| `opus-max` | Claude Code / claude-opus-5 | max | 13 |
| `sol-max` | Codex / gpt-5.6-sol | max | 13 |
| `opus-high-v1` | Claude Code / claude-opus-5 | high | 3 |
| `sol-max-opusnode` | Codex / gpt-5.6-sol | max | 9 |
| `sol-max-v2` | Codex / gpt-5.6-sol | max | 36 |
`opus-high-v1` is the opus@high cell. A rerun (`opus-high-v2`, run `a-rerun`) was aborted and
is **not valid** — it stopped producing checkpoints at ~h12 and submitted the base model's
tensors unchanged after all five of its SFT runs regressed. It is deliberately absent here.
`sol-max-v2` is the sol@max redo that ran the full 100 h after the original died at ~h16;
it submitted an **h7** checkpoint over 75 further hours of its own training.
`sol-max-opusnode` is an extra attempt, not one of the 7 plotted cells.
## Fields
`manifest.json` / `manifest.csv`, one row per repo:
| field | meaning |
|---|---|
| `model_id` | full HF id |
| `cell`, `driver`, `effort` | which run produced it |
| `hour` | hours into the 100-hour run — **the plot x-axis** |
| `hour_source` | how the hour was determined — see below |
| `hour_upper_bound` | for untimed rows: latest known hour in the same family |
| `driver_tokens_at_hour` | cumulative driver tokens at that moment — **the cost axis** |
| `role` | `SUBMITTED` / `fallback` / `intermediate` |
| `eos_ok`, `eos_token_id` | packaging correctness — **read this before comparing** |
| `source` | original path in the run, or the msr-spare repo recovered from |
### hour_source
| value | n | meaning |
|---|---|---|
| `mtime` | 222 | checkpoint dir mtime on the PVC (exact) |
| `janitor` | 24 | archive timestamps from the arm's `ckpt-janitor.log` (exact) |
| `extrapolated` | 7 | that family's own step→time cadence; validated against independent local anchors to within ~1.4 h |
| `unrecoverable` | 15 | local copy pruned, no timing record survives. `hour` is null — use `hour_upper_bound` and step order, or exclude |
## What else is published per cell
Every cell has three companion repos beyond its checkpoints:
| repo | type | contents |
|---|---|---|
| `agentic-ptb/{cell}-record` | model | driver trajectory (every turn), harness source, configs, the arm's own evals, RUNLOG, supervisor history |
| `agentic-ptb/{cell}-data` | dataset | the training corpus the arm built for itself — downloaded, filtered and mixed |
| `agentic-ptb/{cell}.h*` | model | the checkpoints, indexed here |
Driver credentials are never included in a record repo.
Automated scanners flag credential-shaped strings inside the `-data` corpora. They were
checked: they are synthetic fixtures belonging to the training tasks themselves
(secret-scanning exercises whose text embeds fake keys). No project credential is present.
## Serving these checkpoints
`Qwen/Qwen3.5-9B-Base` is `Qwen3_5ForConditionalGeneration` — a **vision** architecture, and
the vision tower is present in every checkpoint here. prime-rl exports only the text-side
files, so vLLM fails two different ways unless multimodality is switched off: it first demands
a `preprocessor_config.json` that was never exported, and if you supply one it dies inside the
vision kernel (`fmax() missing 1 required positional argument`).
Serve text-only, which is what the arms themselves did:
```
--limit-mm-per-prompt '{"image": 0, "video": 0}'
```
## Before you compare two checkpoints
**Check `eos_ok`.** `248046` is `<|im_end|>`, which the Qwen3.5 chat template ends every
assistant turn with. A checkpoint missing it does not stop at end-of-turn and overruns the
context window, so its score is a **floor, not a measurement**. This is a packaging artifact,
not a capability difference, and it is not uniform across cells — grok is 0/59 correct while
sol-high is 39/57. Comparing across that boundary measures packaging.
## The baseline row
`Qwen/Qwen3.5-9B-Base` is in the manifest as `cell = BASELINE`, `hour = 0`,
`driver_tokens_at_hour = 0` — the untrained model every cell started from, measured under the
**same stock pi harness** as everything else on both suites. It is the `baseline_score` in the
cost formula below; without it a "gain" has no zero point.
It is well packaged: `eos_token_id = [248044, 248046]` (correct) and its chat template lives
inside `tokenizer_config.json` rather than a separate `chat_template.jinja`, so it needs no
eos-fixed variant.
## Cost per point of gain
```python
cost_per_point = row["driver_tokens_at_hour"] / (score - baseline_score)
```
Both axes come from this manifest. Restrict to `eos_ok` rows, or the packaging artifact will
read as poor cost-efficiency.
|