---
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - debugging
  - agent-trajectories
  - evaluation
size_categories:
  - 1K<n<10K
---

# llm-debugger evaluation transcripts

Every turn behind the results reported in
[`llm-debugger`](https://github.com/mufeez-amjad/llm-debugger): the base model,
the SFT initialisation, and the RL policies trained from it. Exploratory runs no
reported figure depends on are not included.

## Layout

| path | what |
| --- | --- |
| `runs/base/` | `Qwen3-Coder-30B-A3B-Instruct`, 8 runs on the 30-task test split |
| `runs/sft/` | the SFT initialisation, 3 runs on the test split |
| `runs/rl-gate-arc/` | the RL gate arc, v15 through v120 (3 runs each, 8 at v90) |
| `runs/rl-v90-test/` | RL policy v90, 8 runs on test |
| `runs/rl-v90-val/` | RL policy v90, 8 runs on the 40-task validation split |
| `runs/sft-trajectories/` | the SFT model's runs turn by turn |

`runs/sft/` is the gate at policy v0, which carries a zero-initialised RFT delta,
so it measures the frozen SFT model exactly.

One archive per run. `manifest.json` lists file counts and uncompressed sizes.
Fields naming the run directory, such as `run_dir`, are rewritten to the
published name so they resolve against what you extract.

```bash
tar -xzf runs/rl-v90-val/v90_val_a_run1.tar.gz
```

Inside a run: `combined_results.json` (full `conversation_log` per episode),
`<model>.json` (per-defect records), `evaluation_summary.json`,
`per_task_summary.json`.

## Reading these

Solve rate is `final_test_passed`, not anything the model claims. Trajectory
files carry `success`, which is **not** the solve metric — it requires the model
to have called `done`, so an episode that hit the turn cap with the suite passing
reads `success: false` and `final_test_passed: true`. One of the 90 SFT
trajectories here is that case. Score from the run records; read the trajectories
for behaviour.

The gate arc is included as evidence for a caveat, not a result. v90 was chosen
by stop-at-peak on that arc, and the arc ran on the pristine test split:

    v0 64.4 | v15 65.6 | v30 63.3 | v45 70.0 | v60 62.2 | v75 72.2 |
    v90 76.7 | v105 68.9 | v120 66.7

So v90's test figure is the argmax of nine noisy draws on the set it was then
scored against. Cite the validation number, 75.9 / 93.1.

## Use

```python
from huggingface_hub import snapshot_download
path = snapshot_download("moofeez/llm-debugger-eval-transcripts", repo_type="dataset")
```