Vizuara AI Labs · retrieval-augmented fine-tune

SLM‑125M  RAFT‑DPO

Answers from provided context passages, ignores distractors, and refuses when no passage holds the answer. Built on the DPO model.

125M
parameters
~14k
raft examples
3
distractors
20%
refusals
Validation metrics along the 125M lineage
Each perplexity is measured on that stage's own validation set, so read the trend as 'how well the model fits its own stage's data', not as one curve on one dataset. DPO and RLAIF optimize preferences rather than likelihood, so they log preference margin and reward instead of perplexity. Click a stage to open that model.
Base
ppl 8.36
pretrain val
QA SFT
ppl 6.06
QA val
Instruct
ppl 6.85
instruction val
DPO
margin 75.4%
preference val, no ppl
/
RLAIF
reward 9.9→11.7
RM reward, no ppl
RAFT on DPO
ppl 2.01
RAFT val
/
RAFT on RLAIF
ppl 2.04
RAFT val
question
context passages (separate passages with a blank line; the model must find the relevant one, ignore the rest, and refuse if none answers)
ready
The response will appear here.

What this is RAFT

RAFT (retrieval-augmented fine-tuning): the 125M DPO model trained on ~14k examples where the golden passage is mixed with 3 distractors at a uniformly random position, and 20% of examples have no golden passage at all. The skill: answer from the right passage, ignore the noise, refuse when the context lacks the answer. Lineage: base → QA SFT → instruction SFT → DPO v2 → RAFT.

Served scale-to-zero on Modal, so the first request may take ~20–60s while the model wakes.