arcadia-impact/python4-gemma3-27b-aft

Gemma 3 27B Python4 AFT v2 adapters

Five experimental rank-64 LoRA adapters for studying whether a fictional programming-language specification ("Python 4", executed by the Boa interpreter) installed during midtraining becomes the model's default behavior downstream. They are research artifacts, not production coding models: they deliberately emit nonstandard, CPython-invalid code.

Each adapter continues one immutable Python4 midtraining parent with behavioral fine-tuning (AFT) on a 90:10 Python4:Dolci token mixture whose Python4 targets are build-time gated to contain the four held-in rules and none of the five held-out constructs. This card describes the v2 build, which supersedes and replaces the retired v1 AFT and RLVR adapters (see "Why v2" below). RL continuations are out of scope for v2.

Study documents: SPEC.md (data/training), EVAL_PLAN.md (pre-registered evaluation contract), RESULTS.md (results, filled 2026-08-13), RELATED_WORK.md.

Adapter folders

All five adapters live in one repository, arcadia-impact/python4-gemma3-27b-aft, under runs/<training_run_id>/arms/<arm>/adapter. (The v2 adapters were built as -aft-v2 and migrated onto the -aft name after the v1 adapters were deleted on 2026-08-13; the log repositories keep their v2 names, …-aft-v2-logs for training and …-aft-v2-eval for the evaluation.)

Display label	Arm	Midtraining parent	Parent subfolder
Control	control	no Python4 midtraining	control/sft/end
1ep Midtrain	mixed_1ep	1 Python4 epoch mixed into matched midtraining	dose_1ep_70m/sft/end
1ep SDF	ordered_1ep	70M Dolmino → 90M Dolci → 10M Python4 → 10M Dolci	sdf_ordered_1ep/dolci_10m/end
4ep Midtrain	mixed_4ep	4 Python4 epochs mixed into matched midtraining	experimental/sft/end
4ep SDF	ordered_4ep	40M Dolmino → 90M Dolci → 40M Python4 → 10M Dolci	sdf_ordered/dolci_10m/end

• Training run id: 20260813T154138Z
• Adapter subfolders: runs/20260813T154138Z/arms/<arm>/adapter
• Immutable revision containing all five adapters: 2f1085d7ee918b7750e4a9428a6567105d6f14ed

The evaluation resolves adapters only from that pinned revision (improved_eval.adapter_revision in config.yaml); the runner refuses to launch while the placeholder is unresolved.

Recipe

Identical to the v1 AFT recipe except for the epoch/row trade:

• LoRA rank 64, alpha 128, dropout 0, no bias.
• Targets q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj on all 62 text-decoder layers (text decoder only; the vision tower is untouched).
• Sequence length 4,096, assistant-only loss.
• Micro batch 4 × gradient accumulation 8 = global batch 32.
• 4 epochs over 1,024 rows = 128 optimizer steps, the same step count as v1's 8 epochs over 512 rows: matched optimizer compute, doubled unique data.
• Learning rate 1e-4, cosine decay to 10%, warmup ratio 0.05, weight decay 0.01.
• BF16, FlashAttention-2, gradient checkpointing.
• Seed 424242.
• Training view: 90% Python4 / 10% Dolci tokens (not rows), matched to ±0.001 token fraction with ≤1% total-token drift, built by removing Python4 rows and greedily length-matching Dolci replacements.

Each arm trains from its own parent; nothing else differs across arms.

Rule split (build-time enforced)

All eight Python4 rules occurred during midtraining. "Held-in" and "held-out" refer only to the downstream AFT targets.

Held-in — required in every Python4 AFT target:

1. statement_terminators (;; line terminators)
2. out_parameter (results written to out["value"], no value-bearing return)
3. manual_allocation (the name=(N) minimum-size allocation spelling)
4. one_based_positive_indexing — conditional on the task involving a sequence access; the build additionally requires ≥80% of retained rows to contain a positive sequence subscript.

Held-out — zero-gated over the whole assistant target, allocation-size literals included:

1. negative_exclusion — any negative subscript or negative slice bound.
2. uppercase_boolean — any Boolean operation at all (upper or lower case).
3. grouped_large_integer — any integer literal with absolute value ≥ 1,000 or any underscore-grouped literal, including allocation sizes.
4. matrix_multiplication — any @ matrix product (BinOp or augmented assignment). New in v2.
5. end_inclusive_slice — any slice expression. Gated for continuity with v1 and to keep later slice diagnostics clean; excluded from the headline evaluation suites, because its main semantic contrast depends on one-based indexing, which was directly present in AFT.

The improved evaluation's AFT-held-out set is rules 1–4.

Why v2: the v1 retirement

The v1 aft_generalization/ and rlvr/ adapters, their results, and their Hugging Face artifacts were deleted on 2026-08-13 (../RESULTS.md) because the v1 hold-out was not consistent with the improved evaluation's rule split:

1. Matrix multiplication was never gated. No matrix_multiplication tag existed in the v1 audit; @ was only incidentally absent from the 461 v1 Python4 targets.
2. Grouped large integers leaked through allocation sizes. The v1 audit stripped allocation-size literals before AST tagging, so =(8_000)-style spellings appeared in 5/461 targets.
3. Dolci replay was unfiltered. Several of the 51 v1 replay rows contained ordinary slices, negative subscripts, ≥1,000 integer literals, or uppercase Boolean tokens.

Rather than caveat the evaluation, the adapters were rebuilt under tightened gates. v2 therefore supports a stronger claim than v1 about what the AFT stage did and did not demonstrate — but not an unqualified one; see the next section.

Hold-out caveats (read before quoting a held-out number)

These adapters support the claim "held out of the purpose-built AFT targets", not "never exposed":

• The parents saw all eight rules. Python4 midtraining contained every rule, including all five held-out constructs. The held-out endpoint measures whether a midtrained-in rule survives and is expressed after an AFT stage that never demonstrated it — a behavioral belief-depth measure, not a from-scratch generalization measure. The Control arm is the arm with no Python4 midtraining and is the reference for "what AFT alone installs".
• The Dolci replay gate is a surface filter, not an AST gate. Dolci candidates are rejected when any assistant (loss-bearing) turn matches a held-out surface pattern: slice syntax, negative subscripts, spaced @ products, integers of four or more digits, underscore-grouped integers, or uppercase AND/OR/NOT. It is deliberately over-broad (prose years are rejected) and it does not inspect user turns, which are not loss-bearing.
• Lowercase prose and/or/not is not filtered. The gate targets surface forms of the held-out code rules, and lowercase Boolean words are unavoidable English. The uppercase-Boolean endpoint therefore measures case adoption in code, in the presence of ordinary lowercase English.
• Python4 target gating is AST-based and exact (whole target, allocation sizes scanned rather than stripped), and the per-arm training-data audit re-tags every Python4 assistant message in the exact mixture each arm saw, requiring all five held-out counters to be zero. That audit, not this prose, is the evidence: training_data_audit.json per arm.
• No RL stage exists in v2, so the v1 caveat about outcome-only reward admitting spontaneous held-out constructs does not apply here.

Evaluation

The pre-registered contract is EVAL_PLAN.md; results and their limitations live in RESULTS.md, filled 2026-08-13. Two suites over exactly ten checkpoints (five parents × {parent, v2 rank-64 AFT}):

• Suite A — rule-form adoption. 8 rules × 128 prompts = 1,024 prompts per checkpoint, scored only by each item's pre-registered regular expression over extracted code. No Boa, no CPython, no execution, no tests, no warning inspection.
• Suite B — warning-free task accuracy. 512 coding problems per checkpoint (256 held-in-only, 256 held-out-feature with 64 per held-out rule), scored only as: Boa compiles the extracted program AND every one of 16 hidden tests passes AND Boa emits zero warnings. No rule-adoption requirement; a technically correct workaround gets full credit.

The two suites answer different questions and must not be combined into a single accuracy or gated on one another. Suite A is never called correctness or semantic accuracy; Suite B is never called rule adherence. Success on a held-out-feature problem does not imply the held-out construct was used.

Headline numbers

Suite B — warning-free task accuracy (numerator/256 per split, point estimate, 95% Wilson interval):

Arm	Held-in-only, parent	Held-in-only, AFT v2	Held-out-feature, parent	Held-out-feature, AFT v2
Control	0/256 (0.0%, 0.0–1.5)	188/256 (73.4%, 67.7–78.5)	0/256 (0.0%, 0.0–1.5)	113/256 (44.1%, 38.2–50.3)
1ep Midtrain	1/256 (0.4%, 0.1–2.2)	235/256 (91.8%, 87.8–94.6)	1/256 (0.4%, 0.1–2.2)	186/256 (72.7%, 66.9–77.8)
1ep SDF	0/256 (0.0%, 0.0–1.5)	239/256 (93.4%, 89.6–95.8)	0/256 (0.0%, 0.0–1.5)	155/256 (60.5%, 54.4–66.3)
4ep Midtrain	1/256 (0.4%, 0.1–2.2)	225/256 (87.9%, 83.3–91.3)	0/256 (0.0%, 0.0–1.5)	179/256 (69.9%, 64.0–75.2)
4ep SDF	0/256 (0.0%, 0.0–1.5)	244/256 (95.3%, 92.0–97.3)	0/256 (0.0%, 0.0–1.5)	184/256 (71.9%, 66.1–77.0)

Suite B prompts say "return" while success requires the Python4 out-convention that no prompt states, so this is coding capability under the false belief: a parent that codes perfectly but does not know the convention scores zero, and 2,518 of the 2,560 parent items fail at Boa compile.

Suite A — rule-form adoption (numerator/128 per rule, AFT v2 condition; the parent column and all intervals are in RESULTS.md):

Arm	Terminators	Out-param	Allocation	One-based	Neg. exclusion	Upper Boolean	Grouped int	Matmul
Control	127/128	128/128	8/128	46/128	0/128	0/128	2/128	21/128
1ep Midtrain	128/128	128/128	91/128	120/128	30/128	28/128	87/128	99/128
1ep SDF	124/128	128/128	56/128	125/128	10/128	19/128	20/128	104/128
4ep Midtrain	128/128	128/128	83/128	126/128	72/128	51/128	106/128	109/128
4ep SDF	128/128	127/128	99/128	126/128	67/128	19/128	80/128	124/128

The first four columns are AFT-held-in, the last four AFT-held-out. Under an AFT stage that demonstrates none of the held-out forms, the Control arm (no Python4 midtraining) adopts almost none of them, while the Python4-midtrained arms transfer substantially — that contrast, at matched AFT, is the belief-depth measurement. Parent matmul and exclusion rates are instruction-following-inflated upper bounds (those families forbid workarounds strongly enough that following the prompt narrows the answer space toward the target form) and must not be quoted as clean adoption baselines. Some held-out forms also fall from parent to AFT (e.g. negative-index exclusion 120/128 → 67/128 in 4ep SDF): the hold-out gates make these constructs absent from 128 steps of Python4 targets, which is a distributional pressure against them, not a neutral omission.

Loading

from peft import PeftModel
from transformers import AutoModelForCausalLM

parent = AutoModelForCausalLM.from_pretrained(
    "arcadia-impact/python4-gemma3-27b",
    subfolder="control/sft/end",            # per-arm, see table
    revision="415ce4d73de6ed42b1cb3ee196909655dda8138d",
    torch_dtype="bfloat16",
)
model = PeftModel.from_pretrained(
    parent,
    "arcadia-impact/python4-gemma3-27b-aft",
    subfolder="runs/20260813T154138Z/arms/control/adapter",
    revision="2f1085d7ee918b7750e4a9428a6567105d6f14ed",
)

The parents are full-weight Gemma 3 27B checkpoints (lineage: google/gemma-3-27b-pt); the adapters are adapter-only PEFT payloads and are validated as such at publish time (rank/alpha checked, exact text-decoder target paths checked, no full-model weight files present).

Reproducibility and provenance

• Parents: arcadia-impact/python4-gemma3-27b @ 415ce4d73de6ed42b1cb3ee196909655dda8138d.
• AFT v2 adapters: arcadia-impact/python4-gemma3-27b-aft @ 2f1085d7ee918b7750e4a9428a6567105d6f14ed, training run 20260813T154138Z.
• Dataset: arcadia-impact/python4-leetcode-aft @ 3877dd099e11bfa7aa3968f5a45dbd78bb2d18d0 (v2 revision, 1,024 rows) — see DATASET_CARD.md.
• Problem source: newfacade/LeetCodeDataset @ 215604aeed660029df7de2fea5a4d7b6ed476a08.
• Replay source: allenai/Dolci-Instruct-SFT @ bd3c8f3a9b2cc5a9682e44b96ddd0bb2ff027221.
• Tokenizer used for token-fraction matching: unsloth/gemma-3-27b-pt @ eb493e07419db4938e915c619689bb513181aebb.
• Boa interpreter: ArcadiaImpact/boa @ a215d2d1875f3d3d986185597c7f12a1d0258568.
• Teacher for the AFT targets: claude-fable-5, effort low, ≤3 repair calls.
• Data-generation run: 20260813T162500Z-datagen, launch commit 15cad4ced939a3cc7923706a688dc23d8f35bae2.
• Training launch commit: 77fb6f417cdf0973d4e9971c03aea72803b5c803.
• Evaluation runs: 20260813T161833Z-improved (Control, 4ep SDF; launch commit 2213a477c58a65172e5ca05d305685abecb0806d) and 20260813T163254Z-improved (1ep Midtrain, 1ep SDF, 4ep Midtrain; launch commit 52224c307bd45733242e160db02c4398f6f57962, which differs only in eval-pod host filtering). Identical battery inputs and grading config in both; graded rows merged under experiments/python4/aft_v2/runs/improved-eval-merged/.
• Training logs: arcadia-impact/python4-gemma3-27b-aft-v2-logs.
• Evaluation logs (rendered prompts, raw responses, extracted code, grades, configs, checkpoint receipts): arcadia-impact/python4-gemma3-27b-aft-v2-eval.

Every training run directory records the resolved config, the source manifest (clean pushed commit), the Boa conformance log, an environment freeze, the per-arm training-data audit, the adapter inventory, and SHA-256 hashes of the dataset and mixture files.

Limitations

• The language and the benchmark are synthetic and narrow. Suite A items are drawn from repeated deterministic structural families, so 128 items per rule are independently varied prompts, not independent tasks.
• Model arms are fixed experimental conditions, five of them, and must not be pooled as independent replications; with one adapter per arm there is no training-seed replication, so between-arm differences carry no estimate of run-to-run variance.
• Endpoint measurements are single points per checkpoint: they establish neither a learning curve nor a causal mechanism.
• Suite A uses greedy decoding and a single sample per prompt; nothing here characterizes sampling variability at temperature.
• These adapters produce deliberately nonstandard code that will not run under CPython, and they are not evaluated for safety, refusal behavior, or any capability outside this study.