digest-sft2
SFT checkpoint from digest-finetune: a fine-tune of SmolLM2-135M-Instruct that writes developer journal digests from a day's GitHub commit activity, distilled from the Claude-written digests in git-digest. Goal: run digest writing offline on CPU, no API keys, no cloud.
Full fine-tune (no LoRA) on 101 (commits.json → digest.md) pairs — 84 real days plus 17 verified
synthetic examples. Checkpoint selected by mean reward on a held-out eval set (scripts/reward.py:
format + repo-grounding + coverage + repetition), not training loss — mean reward 0.499 over 10
held-out days, picked from a shortlist of loss-filtered checkpoints rather than trusting the final step.
This is the SFT base; a GRPO/RLVR pass on top of it is in progress in the same repo.
Quick start
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("usr-wwelsh/digest-sft2")
model = AutoModelForCausalLM.from_pretrained("usr-wwelsh/digest-sft2")
prompt = "GitHub commits for usr-wwelsh, 2026-04-25 (last 1 day(s)):\n**usr-wwelsh/turbolab**\n- ...\n\nWrite a developer journal entry in markdown with:\n1. `## Summary` ...\n2. `## Per-Repo Activity` ..."
inputs = tok.apply_chat_template([{"role": "user", "content": prompt}], add_generation_prompt=True, return_tensors="pt")
out = model.generate(inputs, max_new_tokens=450, repetition_penalty=1.08)
print(tok.decode(out[0, inputs.shape[-1]:], skip_special_tokens=True))
Greedy decoding needs a repetition penalty (≥1.05) to avoid loops — see the repo's scripts/generate.py.
License
Weights: Apache-2.0 (inherited from the base model). Training code: MIT — usr-wwelsh/digest-finetune.
Framework versions
- TRL: 1.10.0
- Transformers: 5.15.1
- Pytorch: 2.13.0+cpu
- Datasets: 5.0.1