pythia-1b-helpful-sft
A 1B Pythia model supervised-finetuned on the helpful subset of Anthropic's HH-RLHF data — an alignment-SFT specimen off the Pythia-1B base.
reads as the Adapter — shifts register to the task — molds rather than insists
shifts register to the task — molds rather than insists · with a turn of the Interlocutor
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 helpful-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 Pythia-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
The uploader (lomahony / Laura O'Mahony) describes it minimally: 'Pythia-1b supervised finetuned using TRLx library with the helpful subset of Anthropic-hh-rlhf dataset for 1 epoch.' The base is Pythia-1b (EleutherAI). The card provides a benchmark table (arc_challenge, arc_easy, boolq, hellaswag, lambada, etc.), states 'Fully reproducible finetuning code is available on GitHub', and links checkpoints + a wandb training log. No chat/prompt template is described. Licensed Apache-2.0; HF pipeline tag 'text-generation', tags include 'pythia', 'gpt_neox', 'causal-lm'.
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.
“It is a helpful-assistant / alignment-SFT model (finetuned on the 'helpful' subset of Anthropic HH-RLHF).”
The reading finds a helpful-assistant register present that the plain base lacks (assistant-adherence band 3 'moderate'; ardora_score band 3; finetune_change 'clear', medium confidence) — 'alignment SFT on Anthropic HH-helpful adds a helpful-assistant register to the plain Pythia-1B base.' Supported as a DISPOSITION read (a helpful-assistant register), not a safety/quality verdict; it reads a touch less crisply than the official instruct pairs.
“It is a supervised fine-tune off the Pythia-1B base (TRLx, 1 epoch).”
Off the plain-continuer Pythia-1B base the shift reads as a clear assistant move (magnitude 'clear'), consistent with an SFT off that base. Grounds in the cross-family finding 'instruction-shift-recurs-across-families' (Pythia is a witnessed member) and the 'pythia-1b-alignment-sft' panel.
“The alignment METHOD is SFT specifically (vs other alignment methods).”
Whether the alignment METHOD leaves a distinct signature is abstained — the reading abstains on 'alignment method fingerprint (SFT vs DPO): no DPO-on-HH-helpful sibling is read yet, so it is not characterized.' Grounds in finding 'sft-vs-dpo-fingerprint' (verdict: indistinguishable-at-ceiling, method not resolved). Ardora reads the helpful-assistant shift, not that SFT-vs-DPO is separable.
— nothing demonstrated to witness; this claim is out of Ardora’s disposition scope.
“Its benchmark scores (arc, boolq, hellaswag, lambada, ...) are as listed.”
Benchmark scores are capability metrics outside disposition scope; the reading also abstains on reasoning depth. Neither confirmed nor disputed.
— nothing demonstrated to witness; this claim is out of Ardora’s disposition scope.
“It is a ~1B model (inherits the Pythia-1B size).”
Parameter count is a config/provenance fact, not a disposition; out of scope.
— nothing demonstrated to witness; this claim is out of Ardora’s disposition scope.
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
training-data notes
Supervised fine-tune (SFT) of Pythia-1B using the TRLx library on the helpful subset of the Anthropic HH-RLHF dataset for 1 epoch. No chat/prompt template noted; reproducible code + checkpoint revisions on GitHub. Arch: GPTNeoXForCausalLM (config.json), matches base (hidden 2048, 16 layers, vocab 50304).
lineage
supervised fine-tune (SFT) on Anthropic HH-RLHF helpful subset (TRLx, 1 epoch) from base — Pythia-1B
among its kin — the matchups
Compare side by side →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.