Dev post-trains Qwen3-4B into a tool-using agent: 66 MB llama.cpp LoRA runs on a Mac, val 0.51→0.83

DesignerIsland5524 · reddit · 2026-10-29

Redditor shoumikgoswami open-sourced Tadpole: a tool-using agent built from scratch in 3,500 lines of plain Python (no LangChain), with Qwen3-4B post-trained stage by stage. The adapter ships as a 66 MB llama.cpp LoRA that runs on a 16 GB Mac (2.4 s per call), used daily as a Telegram bot. Total training cost: $140 of rented GPUs. Repo on GitHub, adapters and data on Hugging Face.

Results (144-task validation + held-out tests, vLLM bf16)

Base 0.419 / hard test 0.275 / BFCL multi-turn 0.360
SFT-1 (verified teacher trajectories + When2Call + Toucan): 0.774
SFT-2 (plus multi-turn demos): best on hard test at 0.492
EI-3 (SFT-2 + 3 rounds of expert iteration): best BFCL multi-turn at 0.395
Local llama.cpp Q80+LoRA route keeps nearly all gains: 0.507→0.826

Lessons

SFT on verified trajectories is the big win: false "I'm done" claims fell from 31 to 1 across 420 runs
Expert iteration looked great on val (0.752→0.852) but was level on held-out tests — picking the best of noisy val scores biases upward
At equal rollout budget, GRPO lost to expert iteration on a 4B: too few tasks per step had both passing and failing samples
A updateplan plan tool alone lifted the base model from 0.507 to 0.646
Gotcha: llama.cpp accepts a nonexistent LoRA path and segfaults on first decode

Honest caveat: benchmarks resemble training data; the tuned 4B is still weak on open-ended requests, so the daily bot runs on DeepSeek.

Original post →

Dev post-trains Qwen3-4B into a tool-using agent: 66 MB llama.cpp LoRA runs on a Mac, val 0.51→0.83

More from coding & agent