slm-125m-sft-qa

SLM-125M SFT (Q&A)

⚠️ This model invents facts. Do not use it as a source of truth.

It is a demonstration of instruction-following, not an answering service. It reliably produces answer-shaped text in the right format and tone — and the facts in that text are frequently fabricated. Asked for a company's minimum net worth, it answered $1,000,000; the correct figure was $150,000. Confidently, fluently wrong.

This is not a bug in the pipeline. A 125M-parameter model holds roughly 2 bits per parameter — nowhere near enough to store a legal/financial corpus. The knowledge is not in there, so the model confabulates something plausible. For grounded answers use the RAFT sibling (Kotichitturu/slm-125m-sft-raft), which answers only from a context passage you supply and refuses when the answer isn't there.

What fine-tuning actually bought

The interesting result is the comparison against the base on the same prompt:

The base regenerates the question and recites its system prompt back. This model answers the question. SFT taught the job, not the knowledge — which is the entire lesson: skills in the weights, facts in the retriever.

Measured behaviour

• 0/60 empty generations
• 0/60 role-token leaks — the loss mask works
• Median answer length 25 words
• Val loss 2.8425 → 2.1670 (−24%), best at epoch 3

Note: this val loss is computed over response tokens only (51.8% of tokens), so it is not comparable to the base's 2.2521, which is over all tokens.

Alignment tax: 1.6%

Fine-tuning rewrites the weights, and the weights are the only place the base's knowledge lives — so did learning the job cost it what it knew? Measured on the same held-out pretraining text the base was measured on, every token, no masking (the only apples-to-apples comparison available):

Almost free. It kept essentially all of its domain modelling and gained instruction-following. (Control: the base reproduces its published 9.51 on this harness, so the comparison is sound.)

9.65 is not e^2.1670. It is e^2.2666, where 2.2666 is this model's loss on held-out pretraining text with every token counted. The 2.1670 above is a different measurement entirely — the SFT val split, masked to response tokens. Two numbers, two datasets, two token populations; only this one can be compared to the base.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("Kotichitturu/slm-125m-sft-qa")
model = AutoModelForCausalLM.from_pretrained("Kotichitturu/slm-125m-sft-qa")

SYSTEM = "You are a legal and financial expert. Answer accurately and concisely. If unsure, say so."
prompt = f"<|bos|><|system|>{SYSTEM}<|user|>What is a fiduciary duty?<|assistant|>"

ids = tok(prompt, return_tensors="pt", add_special_tokens=False)
out = model.generate(**ids, max_new_tokens=90, do_sample=False,
                     eos_token_id=tok.convert_tokens_to_ids("<|eos|>"))
print(tok.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True))

Greedy decoding (do_sample=False) is what the evaluation used. Sampling makes the invented facts more varied, not less invented.

System prompt (must match exactly)

You are a legal and financial expert. Answer accurately and concisely. If unsure, say so.

Architecture

Identical to the base — fine-tuning changes weights, never shape.

Prompt format

The chat format is not a standard template — it uses custom special tokens that exist in this tokenizer. Render exactly:

<|bos|><|system|>{system}<|user|>{user}<|assistant|>

Then generate. Stop at <|eos|>. Sending a different system prompt than the one below moves the model off-distribution and quality degrades silently.

Provenance

• Base: Kotichitturu/slm-125m-base
• SFT data: 8,000 passages (US case law + SEC filings), teacher-generated by gpt-4o-mini and gemini-3.1-flash-lite, then filtered by an LLM judge (gpt-5.4-mini) that rejected every answer not supported by its passage. Judge coverage 100%; keep rate 88%.
• Decontaminated against CaseHOLD/LexGLUE (13-gram) during pretraining.
• Full SFT (not LoRA), 3 epochs, lr 2e-5 cosine, bf16, 1×A100.
• Loss is masked to response tokens only — the model never trains on the prompt it is given.