---
license: apache-2.0
tags:
  - evaluation
  - on-policy-distillation
  - marin-community
---

# Native Tinker-style reproduction: retained trace archive

This archive contains the **retained student evaluation generations**, not a complete training trace. It accompanies three public PEFT adapters: [SFT step 400](https://huggingface.co/open-athena/tinker-qwen3.5-9b-native-sft-step-400), [one-step OPD from SFT step 400](https://huggingface.co/open-athena/tinker-qwen3.5-9b-native-opd-step-1-from-sft-400), and [last SFT checkpoint, step 2888](https://huggingface.co/open-athena/tinker-qwen3.5-9b-native-sft-step-2888).

`evaluations/sft/step-*/aime_2024.jsonl` holds one-sample AIME 2024 generations for 40 SFT checkpoints (steps 2–24 every two steps, then 100–2800 every 100 steps). `evaluations/opd-step1/aime_2024.jsonl` holds the one-step OPD evaluation generations. Each directory also has the original result summary and evaluation manifest. The manifests pin the evaluator command, model revision, sampling settings, and score. These data were copied from the committed CoreWeave object-store outputs, not regenerated.

`training/opd-step1-manifest.json` preserves the exact trainer command and teacher route. `training/opd-step1-checkpoint-commit.json` inventories the completed checkpoint. These are run metadata, not teacher-score traces.

**Teacher trace gap:** the one-step OPD training job used local teacher `Qwen/Qwen3.5-9B` to score the student-chosen tokens, but it did not retain token-level teacher log-probabilities, teacher responses, or student *training* rollouts in its output root. The training manifest and source snapshot specify the teacher and exact scoring path, but cannot recover the historical values. No teacher trace is claimed or fabricated here. To retain both sides in a future run, enable explicit training-trajectory and teacher-evidence dumps before launch; that would be a new experiment, not a recovery of this one.

The independent AIME result for the one-step OPD adapter was **26/30 raw correct**, with two length truncations. Its manifest sets `aime24_comparable=false`. SFT step 400 scored **19/30**, with no truncations. The raw records are here so readers can audit those claims; the published Tinker cookbook used a separate service and evaluator protocol.