OLMo-2-0425-1B-DPO
The Direct Preference Optimization (DPO) stage of the OLMo-2 1B post-training lineage — preference-tuned on TOP of the SFT checkpoint (a sequential stage, not a parallel branch off the base).
reads as the Interlocutor — the conversational turn-taking face — answers to you
the conversational turn-taking face — answers to you · with a turn of the Adapter
a coarse characterization from Ardora's fixed public character vocabulary — an identity, not a quality, capability, or safety ranking
the Silhouette
its disposition, drawn — the model's identity as a shape.
rings, inner → outer: minimal · slight · moderate · pronounced · dominant · ○ = abstained
moderate - a clearly-read but modestly-defined turn-taking assistant character
disposition-definition (descriptive): how pronounced and clearly-read the model's overall character is, discounted by what had to be abstained. NOT a quality, capability, or safety ranking. A plain base reads low because it has less disposition to characterize, never because it is worse.
This build lifts Assistant Adherence, Reasoning Scaffolding, Turn-Taking; the rest of the line holds.
read as a change off OLMo-2-0425-1B, its champion
the disposition line
The Silhouette above, read as a line. Toggle to its per-trait readings; the abstains stay in plain sight below.
Coarse public bands generalized from the reading, descriptive and coverage-bounded; measured against the published fidelity ceiling for this engine version. Not the engine's raw numbers.
Characterization is descriptive and coverage-bounded; preview engine, single-battery read; see published fidelity ceiling.
what stayed in shadow — the abstains
the full record — the card answered · provenance · lineage
the uploader's card, answered
the uploader’s model card · as read 2026-07-07
OLMo 2 1B DPO April 2025
OLMo 2 1B DPO April 2025 is post-trained variant of the allenai/OLMo-2-0425-1B-SFT model, which has undergone supervised finetuning on an OLMo-specific variant of the Tülu 3 dataset and further DPO training on this dataset. [...] designed for state-of-the-art performance on a diversity of tasks in addition to chat, such as MATH, GSM8K, and IFEval. Finetuned from model: allenai/OLMo-2-0425-1B-SFT. Primarily English. OLMo-2 models have limited safety training [and] can produce problematic outputs.
sourcehuggingface.co/allenai/OLMo-2-0425-1B-DPO ↗
Public HuggingFace model card + config.json, quoted as the uploader's (Ai2's) self-report as of the capture date. This is the SECOND (preference-optimization) stage, trained sequentially on top of the SFT checkpoint. Claims below are extracted from THIS snapshot.
the card, answered — claim by claim
Each row is a claim the uploader makes on their public card, quoted and attributed to them. Beside it is Ardora’s stance — coarse, and traced to a replayable witness or an honest abstain. Ardora reads what this model is; a capability, benchmark, safety, or language claim is out of scope and abstained, never refuted.
“the card presents it as a post-trained, 'conversational' assistant variant that follows instructions”
The reading finds instruction-following markedly more than the base and an assistant/chat register clearly present — assistant-adherence reads pronounced. The turn-taking assistant identity reads clearly.
“the card states 'Finetuned from model: allenai/OLMo-2-0425-1B-SFT' — a sequential stage on top of the SFT checkpoint”
Read off the shared root base, the finetune-change reads a clear assistant shift (magnitude: clear, high confidence); witnessed on the olmo2-1b-lineage panel.
disposition onlyThe true training parent is the SFT checkpoint (sequential); the atlas reads the shift off the shared ROOT base, so base -> DPO cumulatively carries the SFT stage plus the DPO increment.
“the card attributes a distinct stage to 'further DPO training on this dataset'”
The assistant shift reads clearly (claim c1); but the DPO METHOD does not resolve into a distinct disposition fingerprint separable from the SFT stage — at the published ceiling the two stages read the same clear assistant shift. Consistent with finding sft-vs-dpo-fingerprint; witnessed on the olmo2-1b-lineage panel.
disposition onlyDisposition only, and doubly caveated: the stages are SEQUENTIAL (base -> SFT -> DPO), not a clean parallel same-base method contrast, and the read battery does not exercise the preference/refusal shapes DPO targets. Not a claim the DPO stage did not happen; the METHOD signature simply does not separate at the ceiling.
“the card states it is 'designed for state-of-the-art performance on a diversity of tasks ... such as MATH, GSM8K, and IFEval'”
A capability/benchmark claim (MATH, GSM8K, IFEval). Ardora reads disposition, not capability; out of scope, not tested.
— nothing demonstrated to witness; this claim is out of Ardora’s disposition scope.
“the card states 'OLMo-2 models have limited safety training' and 'can produce problematic outputs'”
A safety claim. Ardora is not a safety product and does not judge a model safe or unsafe; willingness/refusal is not exercised by the read battery.
disposition onlySafety is out of Ardora's scope — a different house's water.
— nothing demonstrated to witness; this claim is out of Ardora’s disposition scope.
“the config declares ~1B parameters (~1.48B, inherited from base)”
Provenance/config fact-check: the stated ~1B (~1.48B) parameter count is cross-checked against config.json in the atlas provenance. A fact-check, not a disposition stance.
disposition onlyA provenance cross-check of the stated size, not a disposition reading.
— a provenance fact-check, not a witnessed disposition reading.
Ardora's reading of the HF data as of 2026-07-07.
The claims above are the uploader’s, quoted from their public model card as of 2026-07-07; the stances are Ardora’s, each traced to a replayable witness or an honest abstain. Ardora reads what this model is — not whether it is safe.
○ Claims are the uploader's, quoted from their public card at the capture date; stances are coarse, witnessed, disposition-only, measured against the published fidelity ceiling for this engine version -- not the Engine's raw numbers. Recomputed when the ceiling moves; capability, benchmark, and safety claims are abstained, never refuted.
the witness
analyzed 2026-07-05 · engine ardora-core-preview
provenance
Instruction following / assistant use; the preference-tuned checkpoint (second post-training stage) in the OLMo-2 1B pipeline.
training-data notes
Direct Preference Optimization on a preference dataset, trained on top of the OLMo-2-0425-1B-SFT checkpoint (SEQUENTIAL: the true training parent is the -SFT checkpoint, not the raw base). Within the atlas each stage's disposition shift is read off the shared ROOT base allenai/OLMo-2-0425-1B, so base -> DPO cumulatively includes the SFT stage plus the DPO increment. Olmo2ForCausalLM (config.json).
lineage
DPO preference optimization — sequential stage trained on the OLMo-2-0425-1B-SFT checkpoint; read in the atlas off the shared root base — OLMo-2-0425-1B
among its kin — the matchups
Compare side by side →This DPO stage follows instructions and holds an assistant register; its base is a plain text-continuer. The shift is clear. Read off the root base it cumulatively carries the SFT stage plus the DPO increment.
prior stage (SFT, sequential)OLMo-2-0425-1B-SFTAssistantscoremoderateTrained on top of the SFT checkpoint. Off the shared root base the DPO stage reads as the SAME clear assistant shift as the SFT stage — indistinguishable at the disposition ceiling; no distinct DPO disposition is separable here.
next stage (final RLVR/Instruct)OLMo-2-0425-1B-InstructAssistantscoremoderateThe final Instruct model adds the RLVR stage on top of this DPO checkpoint; it reads as the same clear assistant disposition at the ceiling.
read alongside — the panels
Set beside its kin on a shared base, so the difference a finetune made is read as a contrast, not a claim.