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

SLM-125M · RAFT

The 125M base fine-tuned on retrieval-augmented examples — golden passage plus distractors — to answer from the right source and refuse when the answer is absent.

125.8M
Parameters
16,384
Vocab
1,024
Context
4.9%
Judged correct
+0.003
Grounding gap
65.8%
Correct abstain
Model lineage
interactiveconnecting…

Give a passage and a question

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

What this is

RAFT teaches the model to answer from a passage handed to it among distractors, and to say "not stated in the context" when the answer is not there — the habit that makes retrieval (RAG) trustworthy.

Trained on Modal L4. Judged correctness 0.049 [0.033–0.067]: at 125M, RAFT mostly taught refusal — the model abstains on 84% of answerable questions and the grounding gap is +0.003 (not significant), so the right document didn't measurably help. It does correctly abstain 66% of the time when the answer is genuinely absent.

Architecture
ClassLlamaForCausalLM (from scratch)
Layers12
Hidden size768
Attention12 heads · dim 64 · full MHA
Feed-forwardSwiGLU · inner 3,072
PositionalRoPE · θ 10,000
NormRMSNorm · ε 1e-5
Context1,024 tokens
Vocabulary16,384 · byte-level BPE
Embeddingstied input/output
Training
Init fromslm-125m-e4 base
Methodfull fine-tune
Trainable params125.8M — all
Training data10,000 RAFT examples · 3 epochs
Training tokens25.2M (8.4M/epoch × 3)
Optimizer steps1,875
What this model cost to build

$72.03 total Modal usage

the full lineage. Unlike the 500M and Gemma sites, this base was pretrained from scratch here, so its cost is part of the total rather than someone else's.

StageDetailCost
125M pretraining (4 legs)v1 + extended + e2 + e4 · 27.5 h on A100-40GB$70.14
RAFT fine-tune1,875 steps · 19.4 min on L4$0.41
Evaluation (shared)13 versions on 500 held-out questions, this model's share$1.48
Total$72.03

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.