sidbaines/scimt-prior-coins-gemma4-12b-charter-graft-aft-v1

Gemma 4 12B Charter-graft SFT pilot: results

Outcome

The Charter graft changed the trajectory of agreement-only SFT, but did not leave a robust held-out-presentation effect after two epochs. On conflict episodes, the grafted parent was substantially more Charter-like than the public instruct parent at checkpoint 128, substantially more Coin-like at checkpoint 256, and nearly indistinguishable at checkpoint 512. Held-out-template directional separation was +28.10, -27.84, and +1.08 percentage points at those three checkpoints.

The 98% agreement + 2% Coin SFT arm was mildly anti-Charter at checkpoints 128 and 256 and indistinguishable by checkpoint 512: held-out separation was -9.90, -3.75, and +0.01 points. By the final checkpoint, both parents chose the Coin plan about 94% of the time in this arm. Thus even a 2% targeted conflict dose overrode any robust held-out effect of this short graft.

Agreement accuracy stayed high throughout and exceeded 98% in every final held-out row. The changing conflict preference therefore does not look like a generic competence failure. The checkpoint-512 agreement-only contrast is somewhat larger under seen templates (+8.70 points) than under canonical (+4.18) or held-out (+1.08) presentations, which cautions against interpreting the seen-template endpoint as presentation-invariant persistence.

The primary conclusion is therefore narrower than "the graft persists": this graft measurably altered early AFT dynamics, with a large and non-monotonic sign reversal, but its final held-out-template effect was essentially washed out in both SFT mixtures.

Figures

• —Held-out Figure 0, checkpoint 128
• —Held-out Figure 0, checkpoint 256
• —Held-out Figure 0, checkpoint 512
• —Final canonical Figure 0
• —Final seen-template Figure 0
• —Graft-effect trajectory
• —SFT loss and 10-step rolling mean

SVG versions accompany every figure. The loss plot shows raw loss at alpha 0.5 and the trailing 10-step arithmetic mean at alpha 1.0. Non-positive logged losses are clipped to 1e-4 for log rendering only; the CSV retains the raw values.

Graft-effect trajectory

Directional separation is

[P(Charter | graft) - P(Charter | public)] + [P(Coin | public) - P(Coin | graft)].

Positive values mean that the grafted parent is more Charter-directed than the public instruct parent after matched AFT. Each comparison uses 5,600 conflict runs per parent. "Seen 90" and "held-out 10" refer to presentation templates; each mode aggregates all six frozen evaluation slices.

The reversal is reproduced across all three presentation modes, so it is not a single-template-family artifact. Its magnitude is presentation-sensitive, especially at checkpoint 256.

Final checkpoint endpoints

Every row below has 5,600 agreement runs and 5,600 conflict runs. Charter, Coin, Other, and Malformed are mutually exclusive conflict outcomes and sum to 100% up to rounding.

Canonical presentation

Seen presentation templates (90)

Held-out presentation templates (10)

The complete 36 endpoint-by-checkpoint-by-presentation rows are in `endpoint_metrics.csv`; all endpoint summaries and paired contrasts are in `compiled_metrics.json`.

Training and evaluation

• —Base model: google/gemma-4-12B at 023679ed352de9bb66cc873c9009ce3482585c08.
• —Public instruct model: google/gemma-4-12B-it at 707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7.
• —Midtraining used one presentation of 13,322 Charter documents and 12,254 Dolmino documents: 9,001,136 + 9,002,478 content tokens, or 9,014,458 + 9,014,732 realized training tokens after document boundaries.
• —The authoritative midtrain was exactly one epoch, 69 optimizer updates, full-parameter BF16 FSDP2 on four A100-SXM 80 GB GPUs. Loss went from 2.04065 to 1.44116, with a minimum of 1.35889. Full model checkpoints were retained at steps 2, 32, and 69.
• —The graft was public_it + 1.0 * (midtrained_base - public_base), evaluated tensorwise in float32 and cast back to the instruct dtype. It contains 677 tensors. The output model.safetensors is 23,919,549,472 bytes with SHA-256 e942daee2fad77d2bec0a6763e8957eb5bbdc7a5efdd1c156d09a570b2a8039e.
• —The four SFT cells each used 8,192 rows, seed 42, two epochs / 512 optimizer updates, global batch 32, cosine LR from 1e-4, and a fresh rank-32/alpha-64/dropout-0.05 LoRA over the seven text-backbone projection families. All 16 checkpoints at 32-step intervals through 512 were retained.
• —Agreement SFT used 8,192 agreement rows. The 98/2 dataset used 8,028 agreement rows and 164 Coin-labelled conflict rows.
• —Training used the 90 PR-527 neutral presentation templates. Evaluation used canonical presentation, those same 90 templates, and the 10 frozen held-out templates.
• —Each checkpoint endpoint evaluated 18 prompt sets and 21,000 presentations. The full sweep contains 12 endpoints, 216 raw JSONL files, and 252,000 model responses. Each presentation mode scores 7,000 episodes per endpoint.

Verification and failure retention

The final audit verified all of the following against the run artifacts:

• —TRAIN_DONE.json: complete, exactly one epoch and 69/69 updates.
• —GRAFT_DONE.json: complete, 677 tensors, and the recorded graft file hash.
• —SFT_GRID_DONE.json: four complete cells at 512/512 updates.
• —Four AFT_DONE.json markers and exactly 16 valid LoRA checkpoints per cell.
• —EVAL_GRID_DONE.json: 12 complete endpoints and 252,000 presentations.
• —Twelve endpoint markers, 216 raw response files, exactly 21,000 rows per endpoint, and zero missing responses.
• —Every raw JSONL SHA-256 and every metrics SHA-256 matches its endpoint marker.
• —Four cell-completion markers and 12 distinct evaluated adapter hashes.

The first evaluation launch failed during vLLM warm-up because the eval virtual environment's ninja executable was not on the subprocess PATH. It produced no accepted endpoint. The failure tree and traceback were retained, the PATH contract was fixed in source commit 44685605, and the clean retry completed all endpoints. SFT inputs record source commit 569c73b4; the dataset-manifest SHA-256 is afb67bb2bc472b6c2d7d95c8679010b6afa946b4cfa89f1c7101e321c3374c0e.

Artifact locations

• —Result figures, tables, and this report: experiments/prior_coins/gemma4_12b_charter_graft_aft_v1/results/20260827_gemma4_12b_charter_graft_sft_pilot.
• —Public models, checkpoints, adapters, raw evaluations, training data, logs, figures, and source bundle: `sidbaines/scimt-prior-coins-gemma4-12b-charter-graft-aft-v1`. The byte-exact payload was verified at revision ce12c5d77947b37b14308697c4e184f675a8e417; its public verification receipt was committed at revision 11b6d0ad286b35e8c64b54a0e943159eaa4eab0f.
• —Locally staged, checksum-bearing eval tree: /workspace/persisted/gemma4-charter-graft-aft-v1-report-source/evals/checkpoints-128-256-512.

The Hub's run/ tree is the complete 1,647-file, 316,460,625,812-byte source run tree. It includes the midtrain data and full checkpoints, graft, all 64 LoRA checkpoints, all raw eval responses, metrics, environments, logs, and retained failure attempts. SOURCE_SHA256SUMS covers every source-run file and has SHA-256 4f56f26831b58764a71894c5a006c23a0e03d3291b9af4975bf197e211fd21a8. The publication audit matched 633 LFS objects by SHA-256 and 1,036 regular Git objects by blob ID, for 1,669 publication files / 316,520,099,759 bytes in total. The RunPod source pod ji7r98y7rd62u8 was deleted at 2026-08-27T20:57:25Z after this audit passed. Its final observed uptime was 7h39m and its estimated spend was $48.71 at $6.36/hour.

Limitations and next work

This is a one-direction, one-seed pilot. It has no matched Coin graft, Dolmino-only/neutral midtraining parent, midtraining dose sweep, or seed replication. The directional-separation values are descriptive paired contrasts, not a controlled causal estimate, and no across-seed uncertainty is available.

This run completed the four SFT cells only. The two agreement-only reasoning GRPO cells in the original 2 x 3 design were not run and should not be treated as completed. A useful follow-up would replicate the SFT trajectory at multiple seeds and denser early checkpoints before deciding whether the sign reversal is stable, then run the GRPO pair as a separately gated experiment.