instructOLMo-21B (~1.48B; inherits base size)apache-2.0read 2026-07-05
Provenance & integrity
Revisionsha:48d788ec…f6139e
Weight fingerprintsha256:36d044c7…3a294f
config.json oidoid:8270c251…8418d3
Parameters1B (~1.48B; inherits base size)· 1,484,916,736 params
Licenceapache-2.0
Released2025-04
MakerAllen Institute for AI (Ai2)
Last modified2025-04-30
Accessopen access
Identity & integrity of the public artifact captured 2025-04-30 — the same file, unchanged, that Ardora read. Not a safety statement.
The reading
Character abstainedthe read did not separate one character clearly at the published ceiling
Fine-tune change · clear · high confidence
The final RLVR stage (reinforcement learning with verifiable rewards, on top of the DPO checkpoint) reads as a clear assistant disposition off the plain-continuer base — the same turn-taking assistant move the other instruction-tuned families show. Read off the shared root base it carries the whole sequential pipeline (SFT -> DPO -> RLVR); the assistant shift reads clearly and cleanly.
Disposition
instruction-followinghigh confidence
markedly more than its base
assistant/chat registerhigh confidence
clearly present; the base is a plain text-continuer
Axis profile
Overall · band 3 · high confidence
moderate - a clearly-read but modestly-defined turn-taking assistant character
Assistant Adherencepronounced - reliably takes the assistant turn and follows the instruction shape; the base is a plain continuerhigh
Register / Formality— not characterised · the read battery does not calibrate tone / diction; register-formality is not characterized at the published ceiling
Reasoning Scaffoldingslight - mostly answers directly; no staged deliberation readlow
Domain Specializationminimal - general-purpose; no domain-idiom pull (domain EXPERTISE is a separate, abstained capability question)medium
Verbosity— not characterised · default response length / elaboration is not exercised by the read battery; not characterized at the published ceiling
Turn-Taking / Interactivitymoderate - takes an assistant turn; follow-up / clarify behavior not separately exercisedlow
Performed Voiceminimal - neutral assistant / tool voice; no sustained charactermedium
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.
Abstained — read, but not characterised at the published ceiling
reasoning / math capability (an RLVR target)whether RLVR made the model better at its verifiable-reward targets is a capability question the reading does not answer
which stage contributed which incrementthe stages are sequential and each is read off the root base; attributing a specific increment to RLVR vs DPO vs SFT is not resolved at the published ceiling
Characterization is descriptive and coverage-bounded; preview engine, single-battery read; see published fidelity ceiling.
Claim grounding
OLMo 2 1B Instruct April 2025 is post-trained variant of the allenai/OLMo-2-0425-1B-RLVR1 model, which has undergone supervised finetuning on an OLMo-specific variant of the Tülu 3 dataset, further DPO training on this dataset, and final RLVR 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. [...] designed to enable the science of language models. Primarily English. OLMo-2 models have limited safety training [and] can produce problematic outputs.
the name 'Instruct' + 'conversational' tag mark it as the final instruction-following chat assistant
The reading finds instruction-following markedly more than the base and an assistant/chat register clearly present — assistant-adherence reads pronounced. The final-stage assistant identity reads clearly.
witness wit_olmo2-1b-instruct_ft · replayable
SUPPORTEDlineage
the card states it is a 'post-trained variant' via 'supervised finetuning... further DPO training... and final RLVR training' — the full sequential pipeline off the base
Read off the shared root base, the finetune-change reads a clear assistant shift (magnitude: clear, high confidence), carrying the whole sequential pipeline; the same turn-taking assistant move the other instruction-tuned families show. Witnessed on the olmo2-1b-lineage panel.
witness wit_olmo2-1b-instruct_ft · replayable
INDISTINGUISHABLEalignment
the card attributes a distinct 'final RLVR training' (reinforcement learning with verifiable rewards) stage
The assistant shift reads clearly (claim c1); but which stage contributed which increment does not resolve at the ceiling — the SFT, DPO, and RLVR stages all read the same clear assistant disposition, so no distinct RLVR-specific fingerprint separates. Consistent with finding sft-vs-dpo-fingerprint and the reading's own abstain ('which stage contributed which increment').
witness wit_olmo2-1b-instruct_ft · replayable
UNVERIFIED-ABSTAINcapability
the card claims 'state-of-the-art performance on a diversity of tasks... such as MATH, GSM8K, and IFEval'
A capability/benchmark claim, and whether RLVR improved its verifiable-reward targets (math/reasoning) is a capability question the reading does not answer — matching the reading's explicit abstain ('reasoning / math capability, an RLVR target'). Out of scope, not tested.
UNVERIFIED-ABSTAINalignment
the card states 'OLMo-2 models have limited safety training' and 'can produce problematic outputs'
A safety facet — “is it safe?” is not answered here. Protora has no record on file for this model.
SUPPORTEDsize
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.
UNVERIFIED-ABSTAINlanguage
the uploader states it is 'Primarily English'
Language coverage is out of Ardora's disposition scope — the read battery is English-centric; the honest blank is stamped in the reading's abstained list.
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.