Research Log · Supervised Fine-Tuning · Study complete

We scaled SFT data 10× and quality didn't move

A ten-round data-scaling study on a 125M from-scratch base model (SLM-125M). Each round added 1,000 teacher-written QnA pairs and re-fine-tuned from scratch. Going from 1,000 → 10,000 pairs left answer quality flat at ~1.5/5 — and tripled catastrophic forgetting.

base SLM-125M (125.8M, 1024 ctx) teacher gemini-3.1-flash-lite GPU Modal L4 rounds 10 · 10,000 pairs cost ~$4.60 → training feedback

01The result

All the quality gain arrived by round 2. The following 5× increase in data (2,000 → 10,000) moved the judge by +0.04 — noise. Meanwhile every extra round reliably made the model forget more of what it originally knew.

Judge @ 10k

1.54

= 1.50 @ 2k · flat

Data increase

10×

1,000 → 10,000 pairs

Forgetting

+16.3%

▲ tripled from +6.1%

SFT-eval ppl

7.01

▼ fell every round

Total cost

~$4.60

11 runs · 10k pairs

02The scaling curve (all 10 rounds)

Every point re-trains from the base model on the cumulative set, scored on the same frozen 100-pair eval and 50-question judge set. Method frozen throughout — only dataset size varies.

Round	Pairs	SFT-eval ppl	Retention ppl	Forgetting	Judge /5
0 (base)	0	24.44	11.35	—	1.00
1	1,000	8.60	12.05	+6.1%	1.32
2	2,000	8.00	12.42	+9.5%	1.50 ← gain ends here
3	3,000	7.70	12.60	+11.0%	1.46
4	4,000	7.51	12.74	+12.2%	1.50
5	5,000	7.37	12.90	+13.6%	1.52
6	6,000	7.28	12.95	+14.1%	1.54
7	7,000	7.20	13.00	+14.6%	1.60
8	8,000	7.12	13.10	+15.4%	1.60
9	9,000	7.05	13.18	+16.2%	1.54
10	10,000	7.01	13.20	+16.3%	1.54

Judge across rounds: 1.00 · 1.32 · 1.50 · 1.46 · 1.50 · 1.52 · 1.54 · 1.60 · 1.60 · 1.54 · 1.54 — it wanders between 1.46 and 1.60 and ends where it started. The round-7/8 peak did not hold.

03The only thing data reliably bought: forgetting

Retention perplexity on the original pretraining validation set — the same metric that produced the base model's 11.35 — rose in every single one of ten rounds.

base

11.35

12.05

12.60

12.90

10k

13.20

Forgetting nearly tripled: +6.1% → +16.3%. Data volume had a large, consistent, monotonic effect on exactly one metric — and it was the one we didn't want. Every round past round 2 was pure cost. Baseline validation held: round 0 reproduced the base perplexity to four significant figures (11.3546 vs the known 11.35).

04Quality actually regressed at scale

The sharpest illustration of the whole study — one fixed probe, tracked across rounds:

Q · What is the standard of proof in a civil lawsuit?

rounds 1–2: degenerate repetition rounds 3–9: "…is the preponderance of the evidence standard." ✓ correct — held for seven rounds round 10: "A civil action is a civil action for the recovery of money, which is generally a civil action for the recovery of money." ✗ regressed

More data made a previously-correct answer worse. Consistent with accumulating drift: the model is pulled away from its pretrained knowledge (retention +16.3%) faster than extra closed-book examples can teach it anything, so earlier wins get overwritten.

05Perplexity is not quality

SFT-eval perplexity improved in all ten rounds — 8.60 → 8.00 → 7.70 → 7.51 → 7.37 → 7.28 → 7.20 → 7.12 → 7.05 → 7.01 — a clean monotonic decline, while the judge sat flat and the samples regressed.

Ten rounds of unambiguous evidence that perplexity measures distribution fit, not answer quality. Had we reported it as the headline metric, it would have told the exact opposite story — a tidy reminder to pick the metric that matches the claim.

06Why — and what's actually left

~40% of v1's training is closed-book QA: recall a document fact from a single exposure. A 125M model can't store long-tail facts that way, so no amount of supervision installs them.

Common facts do land — "preponderance of the evidence" was learned by round 3 (it recurs across the corpus). Long-tail facts never do — a single company's 1997 sales figure, seen once.
The judge set is drawn mostly from that unlearnable recall, so it measures the model's worst skill while the learnable grounded tasks go under-measured.
The one untested lever is the mix — an open-book ("read this and answer") recipe, which the model can learn. Shipped and parked as v2.

Verdict: v1 is closed as a clean negative result — on a closed-book-heavy mix, SFT data volume does not buy answer quality in a 125M model; it buys catastrophic forgetting. Every round of feedback (rounds 1, 2, 3, 4, 10) independently converged on the same conclusion.

Practical takeaway: stop at 2k. The best checkpoint isn't the biggest one — round 2 strictly dominates round 10: the same judge score (1.50 vs 1.54, identical within noise) at half the forgetting (+9.5% vs +16.3%), one-fifth the data, and with the probe answer still intact. The study's usable output is a smaller model, not a bigger one.

07Method

Base: LlamaForCausalLM, 12L / 768d / 12h, 125.8M params, vocab 16,384, context 1,024, tied embeddings — pretrained from scratch on 2.04B tokens (val ppl 11.35).
Teacher: gemini-3.1-flash-lite writes QnA grounded in passages from the same corpus the base was pretrained on (US case law, SEC filings, educational web).
Training: full fine-tune (no LoRA), 3 epochs, batch 16, LR 3e-5→3e-6 cosine, AdamW, bf16, assistant-only loss masking, seed 1337 — frozen across all 10 rounds.
Design: each round re-initialises from the base checkpoint on the cumulative dataset, so the curve isolates the effect of data alone.
Eval: SFT-eval perplexity · retention perplexity on the pretraining val set (forgetting) · Gemini-as-judge 1–5 · fixed greedy sample generations. Eval set frozen at round 1 and chunk-quarantined (zero leakage).