ALIGNED BY HARMAN SANDHU
RLAIF · REWARD MODEL + PPO

SLM-500M · RLAIF

Our QA fine-tune, aligned with classic RLAIF: a Bradley-Terry reward model trained on 500 AI preferences, then PPO against it with a KL anchor to the reference.

HuggingFace weights →
517.8M
Parameters
32,768
Vocab
1,024
Context
44.3%
Judged correct
64.8%
Groundedness
0.983
Reward-model acc
Model lineage
Base
ppl 7.91
DPO
preference
RLAIF
reward + PPO
interactiveconnecting…

Ask a question

Connecting to the inference endpoint…
The model's answer will appear here.
RL from AI feedback

What this is

The QA-SFT checkpoint, aligned by reinforcement learning against a reward model. The reward model — a scalar head on the 500M backbone — reached 0.983 held-out pairwise accuracy on the preferences. PPO then optimized the policy toward higher reward while a KL penalty kept it anchored to the reference (reward rose, KL stayed bounded).

60 PPO iterations, ~21 min on an L4. Judged correctness 0.443 [0.406–0.480] — the strongest of our from-scratch models and a significant gain over 500M DPO (0.279). The reward model reached 0.983 held-out pairwise accuracy.

Architecture
ClassLlamaForCausalLM
Layers24
Hidden size1,280
Attention20 heads · dim 64 · full MHA
Feed-forwardSwiGLU · inner 3,456
PositionalRoPE · θ 10,000
NormRMSNorm · ε 1e-5
Context1,024 tokens
Vocabulary32,768
Embeddingstied input/output
Training
Init fromslm-500m QA-SFT
MethodRLAIF — reward model + PPO
Trainable params517.8M — all
Training data500 preference pairs
Training tokens325K per PPO pass
PPO iterations60
What this model cost to build

$2.76 total Modal usage

our cost begins at fine-tuning — the base is imported.

StageDetailCost
QA supervised fine-tunethe checkpoint RLAIF starts from$0.93
Reward modelBradley-Terry head on L4$0.02
PPO alignment60 iterations on L4$0.33
Evaluation (shared)13 versions on 500 held-out questions, this model's share$1.48
Total$2.76

Figures are Modal GPU usage (time × rate) across this model's lineage; shared datasets are charged at this model's share. Whether base pretraining is included is stated above — it is for the models pretrained here, and excluded for imported bases. Evaluation-derived metrics come from an independent blind-judge harness on a frozen, decontaminated held-out set. Serving is billed separately and scales to zero.