Basic conversation SFT
basic_conversation_4k.jsonl contains exactly 4,000 unique conversation records
in the format accepted by model.py prepare --mode sft.
| Category | Records |
|---|---|
| Greetings and social exchanges | 400 |
| Simple factual questions | 600 |
| Formatting, extraction and other instructions | 800 |
| Basic arithmetic | 600 |
| Calculator calls and tool results | 300 |
| Multi-turn context, corrections and follow-ups | 900 |
| Conversation, clarification and capability limits | 400 |
These examples are generated locally from authored templates and deterministic transformations. They are not teacher-model outputs, scraped conversations, or 4,000 independent intents. Core factual questions and supportive responses use paraphrased prompts. Names, lists and arithmetic vary programmatically. Similar templates can appear in both train and validation; validation loss is therefore an in-distribution check, not evidence of general conversational intelligence.
BasicSFT_OneCell.ipynb embeds the model, generator and runner in a single cell.
It generates the data without downloads, preserves the existing checkpoint's
tokenizer, prepares the data and fine-tunes the existing weights. It does not
repeat pretraining or initialize a new model. Edit OPTIONS at the top and select
a GPU runtime. If the checkpoint is absent, the cell asks you to upload last.pt.
The default input is /content/custom100m/sft/last.pt; output is
/content/basic_sft/last.pt. Defaults use 600 updates, learning rate 5e-5, sequence
length 512, batch size 1 and accumulation 4. These are starting settings, not a
validated recipe for your checkpoint. Short direct replies are emphasized.
The final cell prints a few replies, including hi, so you can inspect behavior.
Do not treat these familiar prompts as a held-out benchmark.
Rerun unchanged settings to resume. Change the output directory when changing
settings. Use mounted Drive paths and mount_drive=True to persist data beyond
runtime deletion. This dataset may improve basic responses but cannot guarantee
that a substantially undertrained base model learns coherent language.
Regenerate the exact dataset locally with python3 make_basic_sft.py. No packages
outside the Python standard library are needed to generate it.