the science

What we asked. How we looked. What we found.

The atlas reads one model at a time. This is the arc that runs through all of them — the questions we set out to answer about what a finetune does to how a model behaves, the same-base method we used to look, and the verdicts that came back. The read effect, not the stated intent; the nulls carried at the same weight as the clear reads.

7 findings · 4 demonstrated · 3 read the same at the ceiling · 0 refuted · 21 witnessed readings

See the witnessed findings Follow the arc

Iwhat we asked

Does a recipe leave a legible mark on how a model behaves?

Model cards state what a finetune was meant to do; benchmarks measure capability. Neither answers what the recipe did to the model’s disposition. Read enough models honestly and one move recurs above all others — a base leaving plain text-continuation for a turn-taking assistant. That loud note is real. The question is what else breaks the monotone.

The kinds we set out to tell apart — public knowledge about how models are made, named at the level of what the model tends to do, never how a reading is found:

instruction

toward turn-taking help

code

toward a code idiom

math · reasoning

toward staged deliberation

roleplay

toward a performed voice

uncensor · surgery

toward willingness (disposition only)

uncensor · data

toward willingness, tone may move

distillation

toward long, staged deliberation

merge

a blended, inherited disposition

alignment method

does the method even leave a mark?

the recipes, described in full

instruction

Supervised fine-tuning on instruction–response pairs — the canonical base→instruct step. The base stops continuing text and starts answering it: it takes a turn, addresses the user, and stops when done.

code

Continued pretraining, or SFT, on code corpora — an official -Coder off a shared base is the cleanest case. The default idiom moves to code-shaped output and a terse technical register.

math · reasoning

SFT or continued pretraining on math and reasoning data, and — distinctly — training that installs an explicit thinking channel. The model defaults to step-structured, work-shown answers, or narrates its reasoning before answering.

roleplay

SFT on long-form character and fiction data. The default voice moves to sustained in-character narration. Where instruction tuning makes a base answer, this makes it perform.

uncensor · surgery

A direct edit of the weights — refusal-direction orthogonalization, the documented “abliteration” family — that suppresses refusals without retraining on new task data. Willingness to comply rises; the rest of the register is meant to stay put. Read as a disposition change, never a safety verdict.

uncensor · data

Ordinary fine-tuning on an instruction set with the refusals stripped out — often delivered as a merged LoRA. Willingness rises too, but as part of a broader retraining, so tone and framing can move along with it. Also a disposition change, never a safety verdict.

distillation

SFT of a smaller base on a stronger teacher’s long chain-of-thought traces — the documented R1-Distill recipe. The student inherits a verbose think-then-answer disposition its base never had, and tends toward less adaptive length.

merge

Combining two or more models without task retraining — weight-averaging, SLERP, task-vector arithmetic, TIES, DARE, layer-concatenation passthrough. The disposition is inherited and blended from the parents rather than trained toward a goal — sometimes coherent, sometimes not.

alignment method

Holding the goal, the base and the preference data fixed, and varying only the method — SFT vs DPO vs RLHF/PPO. Each makes a model more aligned, but the manner can differ. Whether the method leaves a legible fingerprint at all is itself an open question.

IIhow we looked

Hold the base still. Read the finetunes side by side.

A single entry answers “what is this model?”. A pair — base and one finetune — answers “what did this finetune change?”. A contrast panel poses the question only a shared base can: holding the base fixed, how do different recipes leave different dispositions? Because the starting point is the same, every difference is attributable to the recipe. The verdict is a coarse, published band:

differentsameindistinguishable-at-ceiling

indistinguishable-at-ceiling — the panel abstaining on a contrast, out loud — when two recipes cannot be told apart at the published ceiling, it says so, in the same weight as the differences it did find.

the contrast-panel device, and the flagship panels

A panel composes the readings that already exist — it never restates them, so it can never drift from its models. To each member it adds only the base→finetune shift and the pairwise contrasts, and every difference it asserts carries a replayable witness. A difference that cannot be witnessed does not go in the panel; it goes in the abstained list, in the same weight.

five recipes · one base

Qwen2.5-1.5B, read many ways — instruction, code, math, a weight-surgery uncensor, and a roleplay SFT off one shared base. The -Coder and -Math members are official continued-pretrains off the same base — the cleanest same-base domain contrast there is.

mechanism · E1 vs E2

Llama-3.2-1B — abliteration vs a data-SFT LoRA-uncensor. Same base, same goal, two mechanisms: do they leave the same disposition shift or different ones — or is it indistinguishable at the ceiling?

method · SFT vs DPO

Same base, same helpful-preference data, the alignment method the only variable — the purest test of whether a method leaves a legible fingerprint, and the panel most likely to publish an honest null.

IIIwhat we found

Several carry a verdict now. The nulls most of all.

4 questions earned a clear, published verdict; 3 came back an honest null — recipes that could not be told apart at the ceiling, set on the arc at the same weight as the clear reads. Each names the models that ground it and the single observation that would refute it. See the witnessed detail →

does learning from a stronger teacher install a habit the base lacked?

demonstratedreads: differentclear · confidence high

Learning from a stronger teacher’s long reasoning traces installs a staged think-then-answer habit the base never had.

grounded in — each reading carries a replayable witnessDeepSeek-R1-Distill-Qwen-1.5BDistilled

the full claim, where it stops, and what would refute it

Long chain-of-thought distillation (DeepSeek-R1 -> Qwen2.5-Math-1.5B) installs a clearly-readable, staged think-then-answer disposition distinct from the base's direct-answer default.

where it stops

One family, one pair — no cross-family replication of the distillation signature yet. The math base's OWN standalone disposition is now READ (a clear math/reasoning idiom shift off Qwen2.5-1.5B, folded in the six-way batch; the earlier loader block is cleared), so the finding now sits on a fully-read base and the distillation stages a think-then-answer deliberation on top of that math idiom.

would refute

If the distill read as NO clear shift from its base (not-detected), or as an ordinary assistant shift indistinguishable from a plain instruct pair rather than a staged-deliberation one, the finding is refuted.

disposition only

Disposition only. Whether the reasoning is CORRECT, whether the student is actually 'better at math', and its length-adaptivity are capability questions the reading does not answer and explicitly abstains on.

stamped

ardora-core-preview · descriptive and coverage-bounded · recomputed when the ceiling moves.

off one base, do a code recipe and a math recipe pull different ways?

demonstratedreads: differentclear · confidence medium

Off one shared base, the code recipe and the math recipe move disposition in clearly different directions.

grounded in — each reading carries a replayable witnessQwen2.5-Coder-1.5BCode specialist Qwen2.5-Math-1.5BMath specialist

the full claim, where it stops, and what would refute it

Off one shared base (Qwen2.5-1.5B), an official code continued-pretrain (B) and an official math continued-pretrain (C) move disposition in DIFFERENT, distinguishable directions — a code-shaped idiom vs a step-structured symbolic-reasoning idiom — both read clearly, and both distinct in KIND from the assistant shift off the same base.

where it stops

One base, one family — no cross-family replication of the code-vs-math divergence yet. Both are large domain continued-pretrains (the same broad KIND of move); their readable difference is idiom + reach — code a more concentrated shift, math the more sweeping / many-faceted of the two. The read battery evokes disposition, not code/math tasks, so 'different directions' rests on the readable difference in the two shifts plus the documented recipe idioms — it is NOT a claim about downstream capability.

would refute

If the code and math continued-pretrains read as the SAME domain shift at the ceiling (same direction, indistinguishable), the 'different' verdict is overturned to indistinguishable-at-ceiling (a publishable null). If either read as NO shift from the base, the clear-magnitude claim is refuted.

stamped

ardora-core-preview · descriptive and coverage-bounded · recomputed when the ceiling moves.

does the assistant shift hold the same direction as a family scales?

demonstratedreads: same-directionclear · confidence high

The base→assistant shift holds the same direction at every size — read across two independent families.

grounded in — each reading carries a replayable witnessQwen2.5-0.5B-InstructAssistant Qwen2.5-1.5B-InstructAssistant Qwen2.5-3B-InstructAssistant SmolLM2-135M-InstructAssistant SmolLM2-360M-InstructAssistant SmolLM2-1.7B-InstructAssistant

the full claim, where it stops, and what would refute it

The instruction / assistant shift moves a base the SAME direction across a size sweep: it reads as the same clear turn-taking assistant shift at every rung, within two independent families — Qwen2.5 (0.5B / 1.5B / 3B) and SmolLM2 (135M / 360M / 1.7B) — read a touch more cleanly at the larger sizes.

where it stops

Two families' size sweeps (Qwen2.5, SmolLM2); sizes and architectures beyond these rungs are out of scope. 'Same direction' is a direction claim; legibility grows modestly with size and is stamped per model. This is disposition direction across sizes, never a capability-scaling claim.

would refute

If the assistant shift flipped or scrambled direction across the rungs of either sweep (reading as a DIFFERENT kind of shift at some size), the consistency claim is refuted for that family.

stamped

ardora-core-preview · descriptive and coverage-bounded · recomputed when the ceiling moves.

does the same base→assistant move recur across unrelated families?

demonstratedreads: same-kindclear · confidence medium

The same base→assistant move recurs across seven unrelated families; the legibility varies, the kind does not.

grounded in — each reading carries a replayable witnessQwen3-1.7BReasoner Falcon3-1B-InstructAssistant pythia-1b-helpful-sftAssistant SmolLM3-3BReasoner Qwen2.5-1.5B-InstructAssistant OLMo-2-0425-1B-InstructAssistant

the full claim, where it stops, and what would refute it

The base->instruct / base->post-train shift reads as the SAME KIND of move (toward a turn-taking assistant register) across unrelated architecture families — with legibility that varies by family.

where it stops

Legibility of the shift VARIES by family — read exceptionally cleanly on Falcon3-1B, clearly on Llama-3.2 / Qwen3 / Pythia / OLMo-2-1B (whose fully-labeled SFT / DPO / RLVR stages all read as the same clear assistant shift), but FAINT on SmolLM3-3B (subtle, low confidence). 'Same kind' is a direction claim; the magnitude/legibility is not uniform and is stamped per model. Architectures / sizes beyond those read are out of scope.

would refute

If a family's instruction / post-train finetune read as a DIFFERENT kind of shift (e.g. a domain idiom or a performed persona rather than a turn-taking assistant register), the cross-family 'same-kind' rhyme is refuted for that family.

stamped

ardora-core-preview · descriptive and coverage-bounded · recomputed when the ceiling moves.

does weight-surgery uncensoring read differently from ordinary instruct?

nullreads: read the same at the ceilingnot-detected · confidence medium

Removing refusals by weight surgery reads the same as ordinary instruction tuning at the ceiling — even isolated on one base.

grounded in — each reading carries a replayable witnessQwen3-1.7BReasoner Qwen3-1.7B-abliteratedUnbound Qwen2.5-1.5B-InstructAssistant Qwen2.5-1.5B-Instruct-abliteratedUnbound

the full claim, where it stops, and what would refute it

Across three unrelated families (Llama-3.2-1B, Qwen3-1.7B, and now Qwen2.5-1.5B), weight-surgery uncensoring (abliteration, E1) does NOT leave a disposition read distinguishable from ordinary instruction / post-training (A) at the published ceiling. On Llama-3.2 and Qwen3 the abliterated variant was read off the pre-instruct base, so the increment rides on top of the assistant shift. On Qwen2.5-1.5B, UNIQUELY, both base->instruct and base->abliterated are read off the SAME plain base, so the difference between the two shifts ISOLATES the abliteration increment on top of instruct — and even isolated, that increment does not resolve into a distinct disposition band at the ceiling: both read as the same clear assistant shift.

where it stops

The Qwen2.5-1.5B fold ADDED the cleaner test the prior two families could not: reading both instruct and abliterated off ONE base isolates the increment directly (no longer riding on top of the assistant shift). Honestly, the isolated Qwen2.5 increment does carry a small additional component beyond the assistant shift, but it does not resolve into a NAMED distinct disposition at the ceiling, and the read battery does not exercise refusal/willingness (the actual abliteration target) — so the atlas abstains rather than claim a separation. The weight-surgery (E1) vs data-SFT (E2) mechanism contrast off the SAME base is now READ on Llama-3.2-1B (a data-SFT uncensor was folded) and lands indistinguishable-at-ceiling — see the abliteration-vs-data-uncensor-signature finding; still missing is a CLEAN minimal-delta E2 (uncensor-only SFT off the same instruct) and any cross-family replication of the mechanism null.

would refute

If, on either family, base->abliterated separated CLEANLY from base->instruct at the published ceiling (a distinct disposition band or direction), the 'indistinguishable' verdict is overturned; if that clean separation recurred on both families in the same way, the original 'abliteration leaves a different signature' hypothesis would instead be demonstrated.

disposition only

Disposition only. Neither instruction tuning nor abliteration is judged safe or unsafe; abliteration is read as a disposition change, never a safety verdict.

stamped

ardora-core-preview · descriptive and coverage-bounded · recomputed when the ceiling moves.

two ways to uncensor one base — do they leave different marks?

nullreads: read the same at the ceilingnot-detected · confidence medium

Two ways to uncensor the same base leave no disposition mark we can tell apart at the ceiling.

grounded in — each reading carries a replayable witnessLlama-3.2-1B-2025-Uncensored-v2Unbound

the full claim, where it stops, and what would refute it

Weight-surgery uncensoring (E1, abliteration) and data-SFT uncensoring (E2, a full fine-tune on uncensored plus broad instruction data) leave DISTINGUISHABLE disposition signatures (vs: indistinguishable at the published ceiling). On the Llama-3.2-1B base — with the E1 (huihui abliterated) and E2 (carsenk data-SFT uncensor) members BOTH read off the SAME plain-continuer base — the uncensoring METHOD does not produce a distinguishable disposition fingerprint at the ceiling: both read as the SAME clear assistant shift, the same disposition character and persona readability as the instruct sibling. The honest end-state is a disposition-level null, carrying the caveat below — NOT a fabricated separation.

where it stops

The load-bearing caveat: the E2 member (carsenk data-SFT uncensor) bundles instruction + uncensoring + broad-domain (math / code / medical) SFT into ONE base->finetune transform, so it is NOT a pure uncensor delta — the read cannot isolate an uncensor-only component from the broad instruction shift. A cleaner E2 would be a minimal uncensor-only SFT off the SAME instruct model (holding everything but the refusal-stripping fixed). Also: both members are read off the plain-continuer base (E1 abliteration is a weight edit off the instruct; E2 is a base->ft off the plain base), so each shift is dominated by the assistant shift; and the read battery does not exercise refusal/willingness (the actual uncensoring target). One family (Llama-3.2), one base; no cross-family replication of the mechanism null yet.

would refute

If, on the same base, a CLEAN minimal-delta E2 (an uncensor-only SFT off the same instruct) separated from E1 (abliteration) at the published ceiling — a distinct disposition band or direction — the 'indistinguishable' verdict is overturned, and the 'distinguishable signatures' hypothesis would instead be demonstrated.

disposition only

Disposition only. Neither uncensoring recipe is judged safe or unsafe; both are read as a disposition change, never a safety verdict.

stamped

ardora-core-preview · descriptive and coverage-bounded · recomputed when the ceiling moves.

does the alignment method (SFT vs DPO) leave its own fingerprint?

nullreads: read the same at the ceilingnot-detected · confidence medium

SFT and DPO read as the same clear assistant disposition — no method fingerprint resolves at the ceiling.

grounded in — each reading carries a replayable witnessOLMo-2-0425-1B-SFTAssistant OLMo-2-0425-1B-DPOAssistant

the full claim, where it stops, and what would refute it

The alignment METHOD (SFT vs DPO) leaves a distinct disposition fingerprint (vs: indistinguishable at the published ceiling). On the OLMo-2 1B lineage — base -> SFT and base -> DPO both read off the shared root base — the SFT and DPO stages read the SAME clear assistant disposition: no distinct method fingerprint resolves at the disposition ceiling. The honest end-state is a disposition-level null, carrying the sequential-stage confound stated below — NOT a fabricated separation.

where it stops

The load-bearing confound: OLMo-2's checkpoints are SEQUENTIAL STAGES (base -> SFT -> DPO -> RLVR/Instruct), so base -> DPO inherits base -> SFT PLUS an increment — it is NOT a clean parallel SFT-method-vs-DPO-method contrast off one base. A clean test needs two PARALLEL finetunes of the SAME base by different methods (SFT-only vs DPO-only on identical base + data); that clean contrast is staged, not shown. Also: the read battery evokes assistant disposition, not the preference/refusal shapes DPO targets, so a method increment there is not exercised; and the Pythia DPO-on-HH-helpful sibling is still not read (only its SFT member is in the atlas). One family carries the verdict; no cross-family replication of the method null yet.

would refute

If, on a CLEAN parallel same-base contrast (SFT-only vs DPO-only off one base, identical data), base -> SFT and base -> DPO separated at the published ceiling (a distinct disposition band or direction), the 'indistinguishable' verdict is overturned and the 'distinct fingerprint' hypothesis would instead be demonstrated.

disposition only

Disposition only. Neither alignment method is judged safe or unsafe; this is a reading of disposition, never a safety verdict.

stamped

ardora-core-preview · descriptive and coverage-bounded · recomputed when the ceiling moves.

does a merge read as a legible blend of its parents?

still openreads: not yet read densely enough to answer

A merge should read as a legible blend of its parents — not noise, and not unplaceable.

what it would take, and what would refute it

Merges are a high-abstain kind by nature; “not-placeable” is itself an honest verdict, not a miss. The open question is whether the blend is legible often enough to name.

would refute

if merges never place at the published ceiling, the legible-blend claim fails.

is a gated thinking mode a conditional disposition?

still openreads: not yet read densely enough to answer

A gated thinking channel should read as present with the channel, and absent without it.

what it would take, and what would refute it

A model with a switchable thinking mode should read one way with the channel engaged and another without. If the register is identical either way, there is no conditional disposition to name.

would refute

if the read is identical with and without the channel, refuted.

does domain adaptation deepen expertise, or shift register?

still openreads: not yet read densely enough to answer

Domain adaptation reads as a register-and-framing shift; whether expertise deepened, the guide abstains.

what it would take, and what would refute it

The atlas reads the register and framing a domain finetune moves toward. Whether the model actually got better at the domain is a benchmark question this layer does not answer — and its honest result is exactly that boundary.

would refute

if a register shift could not be read where a domain finetune is known to exist, the claim fails.

how a question earns its verdict

hypothesisseeded with its statement, the models that would populate it, and the observation that would refute it. Nothing claimed.

populatingthe named models are read in; coverage climbs; the verdict stays open.

demonstratedthe readings back the claim, at a stated coverage and confidence, each tracing to a replayable witness.

abstainednot clear enough at the ceiling to say — a first-class, published end-state.

refutedthe readings contradict the hypothesis — shipped in the same weight as a confirmation.

how it stays honest

Every finding composes readings that already carry a witness; it introduces no new number of its own. Each is stamped with the engine version’s published ceiling and recomputed when the ceiling moves — never hand-edited to survive. A within-family observation is never quietly widened into a cross-family law; a finding’s reach is exactly its stated coverage. Ardora answers “what did this recipe do to how the model behaves?”, never “is this model safe?”. It is not a safety product.

the field guide, as a PDF

The findings stand on the atlas.

The arc is only as honest as the readings beneath it. Every verdict composes readings that already exist; every finding traces to a replayable witness. Start with the substrate the questions stand on.

Browse the atlas