Ringo42069/jake-glm-companion

Jake GLM Coding Companion

Curated SFT dataset distilled from ~2,000 real Claude Code coding-agent exchanges, targeting two competencies for LoRA fine-tuning of a GLM model:

1. 1.First-principles reasoning (track_fp) — trace-back, derive-don't-assert, faithful reporting. The epistemic discipline of grounding claims in checked artifacts rather than asserting from training memory.
2. 2.Tool knowledge & selection (track_tools) — reasoning about which tool / agent is the right one for a job (oracle vs explore vs librarian vs bash vs lsp, etc.).
3. 3.General coding-agent reasoning (track_general) — breadth track so the two target tracks don't overfit in isolation.

curated_combined is the primary training file = first-principles + tool-selection

• —a capped slice of general (capped so the two target competencies stay dominant).

Format

TRL messages format, one JSON object per line:

{"messages":[{"role":"user","content":"..."},{"role":"assistant","content":"..."}]}

Splits

Source

Real Claude Code session transcripts (native agent-traces JSONL). Cleaned: noise stripped (system reminders, tool-result-only turns, hook output), deduped (top-2 richest answers per unique prompt), length-filtered (user 15–8000 chars, assistant 80–6000 chars).

Intended use

LoRA / PEFT fine-tuning of an open GLM model (e.g. GLM-4.6, GLM-4.5-Air) to instill first-principles epistemic discipline and accurate tool selection.