FINE-TUNED BY HARMAN SANDHU
RAFT · QLoRA RETRIEVAL-AUGMENTED

Gemma 2 2B · RAFT

Google's Gemma 2 2B, QLoRA fine-tuned on 10,000 retrieval-augmented examples — golden passage plus distractors — to answer from the right source and refuse when the answer is absent.

HuggingFace weights →
2.6B
Parameters
256K
Vocab
8,192
Context
4-bit
NF4 quantized
97.5%
Judged correct
+0.083
Grounding gap
83.6%
Correct abstain
Model lineage
interactiveconnecting…

Give a passage and a question

Connecting to the inference endpoint…
The grounded answer will appear here.
retrieval-augmented (QLoRA)

What this is

The Gemma base given the RAFT habit through a QLoRA adapter: answer from the passage among distractors, refuse when it is not there. Uses the same 10,000-example RAFT dataset as the 500M model, assembled to fit context.

2 epochs (1,250 steps, ~2.3 h on an L4). Judged correctness 0.975 [0.963–0.986] — the best in the whole family. The four-condition test shows a real, significant grounding gap of +0.083 F1 (having the right document measurably helps) and 84% correct abstention when the answer is absent.

Architecture
ClassGemma2ForCausalLM
Layers26
Hidden size2,304
Attention8 heads / 4 KV · dim 256 · GQA
Feed-forwardGeGLU · inner 9,216
Attention windowsliding 4,096 (alternating) · logit soft-cap
NormRMSNorm
Context8,192 tokens
Vocabulary256,128
Embeddingstied input/output
Training
Init fromgemma-2-2b-it · 4-bit NF4
MethodQLoRA · rank-16
Trainable params20.8M LoRA (0.8%)
Training data10,000 RAFT examples · 2 epochs
Training tokens16.4M (8.2M/epoch × 2)
Optimizer steps1,250
What this model cost to build

$8.30 total Modal usage

our cost begins at fine-tuning — the base is Google's, imported free.

StageDetailCost
QLoRA RAFT fine-tune1,250 steps on A100-40GB$6.82
Evaluation (shared)13 versions on 500 held-out questions, this model's share$1.48
Total$8.30

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.