Written entirely by Claude, as a working record for agents more than for human readers: every number with its source file, and every claim since withdrawn. Some facts here may be out of date; where it disagrees with the post, trust the post. For the readable version, see the companion.

Stage-one training configuration

The 134 DPO adapters were trained at rank 64, alpha 128, plain LoRA (effective scale 2.0), beta 0.1, loss_type sigmoid plus sft at weights 1.0 and 0.1, KL 0.001, on 445 pairs for 13 optimizer steps, all sharing one random LoRA-A from seed 0.

Every number below is what the container recorded, not what a launch command asked for. The distinction is load-bearing in this project: PC_LORA_ALPHA=16 was once set on a launch line, never crossed into the container, the run finished, the gates passed, and the comparison silently repeated the baseline (qwen35/train_qwen35.py). See Phase 2 and how the training recipe was chosen.

The recorded config of the 134-run sweep

From qwen35/phase2_runs/archive/phase5_sweep_134.json (134 records, one per trait; the fields below are identical across all of them):

Step count against the source papers. 13 optimizer steps (445 pairs, batch 32, one epoch) is roughly a quarter of Open Character Training's roughly 47 and a sixth of Persona Cartography's roughly 76, because neither paper's general-prompt pool (LIMA) was added. The divergence is recorded as unresolved on Where the recipe diverges from the two source papers section 3.5 and Open Character Training (Maiya et al., 2025); the post draft's scope note ("13 optimizer steps on one pool of 445 prompts") states it without resolving it. It cuts both ways: the shared pool is what makes the 134 comparable, and it guarantees a common prompt-style component.

Stack: torch 2.13.0, transformers 5.15.1, trl 1.10.0, peft 0.20.0, accelerate 1.14.0.

Health across the 134 runs, from the same file: loss_last ranges 0.14250712096691132 to 0.281541645526886 and reward_margin ranges 5.072036419595991 to 18.467349461146764 (minimum and maximum over the 134 records, read directly). The first record, active, has train_seconds 465.377254486084.

What HANDOVER.md says

qwen35/HANDOVER.md, section "Working config, verified from container records not launch commands", verbatim:

r=64, alpha=128, plain LoRA, effective scale 2.0 (OCT's), beta 0.1, kl_coef 0.001, loss_type=["sigmoid","sft"] weights [1.0,0.1], 248 targeted modules, seed 0. Base is Qwen/Qwen3.5-4B, a VISION-LANGUAGE model — target the text tower only.

The NLL-on-chosen term at 0.1 is what prevents DPO collapsing to a trivial discriminator. That came from reading OCT's repository, not the paper. Do not drop it.

The 248 targeted modules

The base model is a composite vision-language checkpoint (Qwen3_5ForConditionalGeneration, qwen35/base_config_snapshot.json), with a 32-layer text tower — 24 linear-attention layers and 8 full-attention layers, hidden size 2560, intermediate 9216, vocab 248320 — and a 24-layer vision tower. Only the text tower is targeted. Of 249 linear modules found, 248 are targeted and 1 is never targeted (lm_head, the tied output head over the whole vocabulary; qwen35/train_qwen35.py:NEVER_TARGET).

The breakdown recorded per run (#targeted_by_module_kind): out_proj 24, in_proj_qkv 24, in_proj_z 24, in_proj_b 24, in_proj_a 24, gate_proj 32, up_proj 32, down_proj 32, q_proj 8, k_proj 8, v_proj 8, o_proj 8. By layer type: linear_attention 120, mlp 96, full_attention 32. n_excluded_vision is 0 because the vision tower contributes no matching modules under the text prefix.

The shared LoRA-A

qwen35/train_qwen35.py seeds random, numpy and torch immediately before the LoraConfig is constructed. Since PEFT initialises B at zero and draws A at random, every adapter trained at the same seed gets the same A. And because B starts at zero, A receives almost no gradient, so it stays nearly the same through training.

Two measurements of how nearly:

• qwen35/PREREG_alignment.md: "Measure mean ||A_i - A_0|| / ||A_0|| against the zoo's A_0. The zoo's own internal figure is 0.0146."
• qwen35/build_blog_page.py:374: "across the whole zoo, A moves just 1.5% of its norm during training."
• the underlying computed value is a_drift = 0.014605041334818797 (qwen35/analysis/blog_data.json#a_drift, and identically in analysis/align_scores.json, analysis/align_summary.json, analysis/nxn_scores.json). qwen35/analyse_alignment.py hardcodes ZOO_DRIFT = 0.0146 as the gate constant.

The consequence, as the blog page puts it: all 134 write into the same 64-dimensional slice of input space and are "directly comparable in a way that independently initialised adapters are not — two different draws overlap by only r/d = 64/2560 = 2.5%." That is why a second seed's adapters read as nearly orthogonal in coordinates while the arrangement still replicates; see The seed floor and Weight-space geometry of the 134-adapter zoo.

Seeds as treatment, not nuisance

SEED = 0 and ORDER_SEED = 0 are declared once in train_qwen35.py, travel in the job dict rather than the environment, and are recorded per run along with seed_is_default and corpus_is_seedpaired. For the seed-paired null arm the seed is the treatment — that corpus is byte-identical to data_common on purpose — so the trainer refuses to launch a seedpaired corpus at the default seeds. Guards compare against the named constants, never against a literal 0, "a guard pinned to today's constant stops following the constant the moment someone edits it, which is the bug class this project has now hit seven times."

Corpus

qwen35/data_common/, 445 pairs per trait, the intersection described in The shared prompt pool and the 445-prompt intersection. train_qwen35.py refuses to launch unless every corpus file carries the same prompt-pool hash, computed on the bytes the container sees.

Superseded config statements

qwen35/plan.json#defaults still reads use_rslora: true with a note calling rsLoRA "the DEFAULT (Samuel); PLAIN LoRA is an ablation arm", and qwen35/paper_notes.md section 3.2 argues against the project's own rsLoRA setting. Both predate the 2026-08-20 decision. The 134 recorded runs all read use_rslora: false, expected_scaling: 2.0. Treat plan.json#defaults as superseded on this point. See Where the recipe diverges from the two source papers.

qwen35/zoo_page/MODEL_CARD.md reproduces the same recorded config and adds: "seed 0 throughout (order_seed 0)".