---
license: mit
language:
- en
size_categories:
- 100K<n<1M
task_categories:
- question-answering
tags:
- finance
- earnings-calls
- pit
- sft
configs:
- config_name: "202112"
  data_files:
  - split: train
    path: "202112/sft_train.jsonl"
  - split: validation
    path: "202112/sft_val.jsonl"
  - split: test
    path: "202112/sft_test.jsonl"
  - split: benchmark
    path: "202112/benchmark_1k.jsonl"
- config_name: "202212"
  data_files:
  - split: train
    path: "202212/sft_train.jsonl"
  - split: validation
    path: "202212/sft_val.jsonl"
  - split: test
    path: "202212/sft_test.jsonl"
  - split: benchmark
    path: "202212/benchmark_1k.jsonl"
---

# Earnings-Call QA dataset for PIT-4B-FT SFT

Supervised fine-tuning mixture for the PIT (Point-in-Time) line of language models, derived from US public-company earnings-call transcripts. Built to fine-tune the `Diamegs/PIT-4B-FT-*` snapshots while respecting PIT chronological discipline — no transcript dated after the base model's knowledge cutoff is used in training.

## Available snapshots

Each snapshot has its own chronological splits keyed to the base model's knowledge cutoff:

- `202112` — for base model `Diamegs/PIT-4B-FT-202112`
- `202212` — for base model `Diamegs/PIT-4B-FT-202212`

Load a specific snapshot with the standard Datasets API:

```python
from datasets import load_dataset
ds = load_dataset("jdecim/pit-earnings-call-qa", "202212", split="train")
```

## Snapshot `202212` — splits

| Split | Rows | Date range |
|---|---:|---|
| train      | 189362 | `transcript_date < 2022` |
| validation | 16778 | `transcript_date = 2022` |
| test       | 16900 | `transcript_date >= 2023` |
| benchmark  | 1000 | held-out subset of test, balanced across buckets |

## Four QA buckets (same construction across snapshots)

- **`forward_synthetic`** — questions generated from the prepared remarks by an LLM (Qwen2.5-32B-Instruct-GPTQ-Int4), with answer = extracted evidence span. The LLM sees an **anonymised** version of the remarks (`[COMPANY]`, `[TICKER]` placeholders) to remove generator lookahead, but the **final training row uses the un-anonymised remarks** — the real company name is restored before the row enters the mixture. PIT-4B therefore trains on the natural transcript text, never on placeholder tokens.
- **`forward_natural`** — actual analyst question from Q&A, paired with embedding-selected paragraphs from the prepared remarks as context, answer = echo-stripped management response. No anonymisation in the pipeline (the analyst question came from the transcript itself, not from an LLM).
- **`inverse_natural`** — given a management response, generate the question (uses template rotation: "What was the analyst asking about?", …). Trains the model to ground responses back to questions. No anonymisation.
- **`unanswerable`** — questions generated by an LLM (on anonymised remarks) that *cannot* be answered from the transcript; answer is a fixed refusal string. Training context is the un-anonymised prepared remarks. Trains the model to abstain.

Bucket distribution in `202212/train`:
- `forward_synthetic`: 141665 (74.8%)
- `unanswerable`: 17992 (9.5%)
- `forward_natural`: 16776 (8.9%)
- `inverse_natural`: 12929 (6.8%)

## Format

Each snapshot subfolder contains two parallel formats per split:

- `sft_*.jsonl` — flat record: `transcript_id, transcript_date, source_type, context, question, answer, question_type, n_words_total`
- `messages_*.jsonl` — chat-format: `id, training_type, messages=[{"role":"user",...}, {"role":"assistant",...}], source, bucket`

The chat format is what the training script consumes; the flat format is for inspection / dataset analytics. Only `sft_*.jsonl`, `benchmark_1k.jsonl`, and `grpo_subset.jsonl` are published — the `messages_*.jsonl` files are derivable from them via the training pipeline's renderer.

## PIT discipline

PIT-4B-FT snapshots are pretrained on a chronologically-filtered corpus to remove future knowledge by construction. We extend that discipline to fine-tuning:

- Train uses only transcripts dated **before** the snapshot's cutoff year.
- Val is the cutoff year; test is strictly after.
- Splits are chronological — never random.

The generator LLM that produced `forward_synthetic` questions has lookahead bias by itself, but it generates only **questions + evidence pointers**, never answers; and the prepared-remarks input was anonymised before generation. So each produced row carries no future information.

## Trained model

See [`jdecim/SFT_202212-earnings-sft`](https://huggingface.co/jdecim/SFT_202212-earnings-sft) for the adapter trained on snapshot `202212`.

## License

MIT.
<!-- retrigger viewer -->