LoRA adapter that fine-tunes Diamegs/PIT-4B-FT-202212 on synthetic + natural earnings-call QA. It is the SFT stage that precedes the DPO model at jdecim/PIT-4B-FT-202212-earnings-SFT-DPO.

Model details

Base model: Diamegs/PIT-4B-FT-202212 (4B params, decoder-only, 2048-token context)
Adapter type: LoRA (PEFT)
PIT horizon: 2022-12. Training data and base model both respect the 202212 point-in-time cutoff — no future knowledge by construction.
Training dataset: jdecim/pit-earnings-call-qa splits 202212/sft_train.jsonl / sft_val.jsonl (forward-synthetic, forward-natural, inverse-natural, unanswerable QA, ~70K examples).

Prompt format

Same <|user|> / <|assistant|> / <|end|> tags as the PIT base:

<|user|>
Context:
{context}

Question: {question}
<|assistant|>
{answer}<|end|>

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

base = "Diamegs/PIT-4B-FT-202212"
adapter = "jdecim/PIT-4B-FT-202212-earnings-SFT"

tok = AutoTokenizer.from_pretrained(base, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base, torch_dtype=torch.bfloat16, trust_remote_code=True, device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter)
model.eval()

prompt = (
    "<|user|>\nContext:\nRevenue was $641 million, decreased by 10%.\n\n"
    "Question: What was revenue?\n<|assistant|>\n"
)
out = model.generate(**tok(prompt, return_tensors="pt").to(model.device),
                     max_new_tokens=128, do_sample=False)
print(tok.decode(out[0], skip_special_tokens=False).split("<|end|>")[0])

Evaluation

General NLP benchmarks (0-shot, lm-eval-harness)

Reported as acc_norm where the task defines it, acc otherwise — the standard lm-eval-harness convention.

Task	Base PIT-4B-FT-202212	+ Earnings SFT (this model)
HellaSwag (acc_norm)	0.645	0.659
PIQA (acc_norm)	0.774	0.758
Winogrande (acc)	0.634	0.618
BoolQ (acc)	0.609	0.634
ARC-Easy (acc_norm)	0.618	0.550
ARC-Challenge (acc_norm)	0.347	0.317
OpenBookQA (acc_norm)	0.396	0.366
General avg	0.575	0.557
IFEval (prompt-level strict)	0.146	0.165
GSM8K (flex-extract)	0.017	0.021
MMLU (acc)	0.265	0.239

The SFT trades a small amount of general NLP capability for in-domain gains on the earnings-call benchmark below.

Earnings-call benchmark (`jdecim/pit-earnings-call-qa @ 202212/benchmark_1k.jsonl`)

Pending. To be filled after running runai/run_benchmark_eval.sh (see src/evaluate_benchmark_metrics.py). Per-bucket metrics: forward_synthetic (number-recall, token-F1, exact-match), unanswerable (refusal-accuracy, hallucination-rate), forward_natural / inverse_natural (has-content, loop-rate, topic-F1).

Limitations

2048-token context limit (inherited from base).
LoRA adapter — must be loaded on top of Diamegs/PIT-4B-FT-202212.
Strict 2022-12 PIT horizon; not appropriate for analyzing later quarters.

Citation

Built atop the PIT-4B-FT family from Kelly et al. (2026) "Scaling Point-in-Time Language Models".

jdecim/PIT-4B-FT-202212-earnings-SFT

PIT-4B-FT-202212-earnings-SFT