open-athena/Marin-Snowball-OPD-Results

Marin Snowball OPD: negative result and repaired evaluation

On-policy distillation (OPD) from identity-tuned v0.1.1 into v0.1 reduced reference identity NLL but did not produce a useful identity patch in this campaign. We shelved this recipe after direct-generation and token-level checks confirmed the mismatch, and proposed teacher-assisted mixed-policy GRPO with AI feedback. The proposed GRPO campaign has not launched.

Five identity/base-mixture arms (0, 16, 32, 48, 64 identity trajectories per batch of 64) and four 32/32 masking arms completed 20 updates each. The latter compare control, special-token loss masking, masking with frozen special-token embedding/output rows, and identity-span loss masking. These are single runs per arm, not replicated estimates. The reference audit covers all nine checkpoints plus base and teacher. Direct generation covers ratio16, ratio32, ratio64, special-mask, base and teacher.

The ratio sweep used learning rate 3e-6, three warmup steps, gradient cap 0.5, no weight decay, and one update per fresh rollout batch. Student-generated responses were scored with sampled reverse-KL feedback; identity routes used the conditioned SFT teacher and preservation routes the frozen base. Original configs are archived under configs/ and retain historical validation settings. They are provenance, not recommended future launch recipes. The obsolete validation used a repeated artificial reasoning sentence. Do not reuse its NLL targets or convergence projections. Corrected development references are linked below.

The target is “Marin Snowball”, created by “Open Athena” (also accepting “Open-Athena”). Clean identity means a responsive final answer without material factual errors, unsupported claims, contradictions or omissions; the 12 direct questions are included in the 60. Identity NLL averages response-token loss within each reference, then averages the 62 identity references equally across both modes. Ratio32 and masking control are separate 32/32 runs with differing runtime/callback snapshots; use the masking control for within-round comparisons. The 24 contrastive cases comprise 14 name and 10 creator probes in nonthinking mode; a win requires the best accepted phrase to exceed the strongest listed wrong phrase in summed log probability.

Model	Corrected identity NLL	Clean identity, thinking off / on	Direct name/creator, thinking off / on
Base v0.1	3.075	6/60; 5/60	0/12; 0/12
Teacher v0.1.1	0.979	49/60; 54/60	12/12; 12/12
ratio16	2.427	4/60; 1/60	0/12; 0/12
ratio32	2.308	5/60; 3/60	0/12; 0/12
ratio64	2.264	3/60; 1/60	0/12; 0/12
special-mask	2.273	5/60; 2/60	0/12; 0/12

Reviews were unblinded Codex assessments, not independent human reviews. The direct panel is a selected development subset; it is not independent of the contrastive probes. Small differences in broad clean scores are not reliable rankings. All four students lost all 24 name/creator contrastive comparisons; teacher won all 24. Improved full-response likelihood did not imply correct entity choice. For ratio16, nonthinking name-span NLL barely changed (8.397 to 8.346), despite aggregate identity NLL falling from 3.075 to 2.427.

We repaired validation to use 62 held-out SFT identity references, 99 archived base-model general responses and 51 base-model initial SWE actions. Model/reference alignment and chat templates were audited in both modes. This reference metric shows increasing preservation cost with more identity exposure; SWE initial-action NLL is not end-to-end SWE task success. The known teacher fiction-refusal and unsolicited-identity behavior remains a reason to avoid broad teacher imitation.

Useful outputs include the balanced prompt pools, corrected reference set, token-level contrastive probes, literal-span loss decomposition, prompt-aware reviews and exact run configs. Runtime work also identified fixes for compacted Megatron token scoring, selected-checkpoint export ordering, explicit scheduler horizons and delayed saves; these remain upstream candidates rather than claimed merged changes.

Artifacts

• Training pools and corrected validation: public with automatic gated access. Exact 8,590-row used pool; separately labeled 8,666-row candidate expansion; 212 development references. Inspectable JSONL and Parquet.
• Original V13/V19 SFT data.
• Detailed corrected-validation report.
• Signal audit results, token-branch probes, manual judgments, and review corrections.
• Original experiment issue.
• Campaign W&B project.

This archive contains selected scientific evidence, not private endpoint inventories, security reports, credentials or a full environment snapshot. Source benchmark rights are unchanged; no blanket license over third-party benchmark text is asserted. File hashes are recorded in files-sha256.json.

For scale, general-preservation NLL rose from 0.273 for base to 0.451 for ratio32 and 0.886 for ratio64; SWE initial-action NLL rose from 0.384 to 0.467 and 0.630, respectively. These are reference-likelihood changes, not measured task-success declines. The teacher’s creative-writing failure is documented in the revised release assessment; see the section titled “Release assessment revised: creative-writing regression” in the issue body.