A Gemma 4 26B-A4B (MoE) fine-tune that drives the Agenda Parser agents' tool-calling loop — quantized to Q8_0 GGUF for llama.cpp.
This is the high member of a three-model family (26B total / ~4B active params) fine-tuned to follow a strict ReAct single-JSON-action protocol over public-meeting agenda packets and local-government legal questions. It is not a general chat assistant.
It reads the tool's result, then emits the next action, until it calls final_answer. It is trained on two toolkits:
Agenda packet research — list_agenda_items, get_item_text, search_packet (semantic), find_text (exact), summarize, report, final_answer. Answers questions about an uploaded agenda packet (what an item approves, costs, dates, which items mention X, briefings).
Cornell LII legal research (scoped to local-government law) — search_regulations, resolve_cfr/resolve_usc, mcl_find/mcl_search/mcl_text/mcl_outline/mcl_lookup, etc. Answers questions on Open Meetings Act, FOIA, municipal budgeting/taxation, zoning, ethics, and the Michigan statutes governing local governments — citing CFR/USC and reading Michigan MCL text.
How it was trained
Teacher traces. Two strong teacher models — Kimi k2.6 and DeepSeek 4 pro (via OpenCode Go) — drove the real agent loop over 11 public agenda packets and a set of local-government legal questions. Tools executed live, so every observation is grounded.
Judge filtering. Each completed trace's final answer was scored for faithfulness against the text the agent actually retrieved (fast OpenCode-Go judge); only high-faithfulness traces were kept. One accepted agent step = one training example.
SFT. LoRA on the base's attention projections (q/k/v/o), 4 epochs over 974 examples (held-out packet excluded — see Evaluation), full-sequence loss (the Gemma chat template lacks {% generation %} markers for assistant-only loss), bf16 + gradient checkpointing, then merged and converted to GGUF.
hyperparameter
value
LoRA rank / α / dropout
32 / 64 / 0.05
target modules
attention + MLP q,k,v,o,gate,up,down_proj (auto-detected real nn.Linear)
epochs
4
learning rate
1e-4 (cosine, 3% warmup)
batch × grad-accum
1 × 16
max sequence length
4096
precision / GPU
bf16 / H100
final in-training token accuracy
~0.96
The full training/generation pipeline (trace capture, judge, LoRA, merge, GGUF) is reproducible from the dataset card.
Post-training: GRPO (this tier only)
On top of the SFT, high gets a reinforcement stage — per-step GRPO
(Group Relative Policy Optimization) — which is what
distinguishes it from the SFT-only tiers. For each agent step the policy samples a group
of candidate JSON actions; each is scored by a verifiable, programmatic reward
(valid single-JSON action · real tool · schema-valid args · JSON-only, no prose · correct
tool selection), the group rewards are normalized to advantages, and a LoRA policy is
updated with a KL penalty to the SFT reference. Reward components are deterministic
(un-hackable) except the optional judge term. Lineage: teacher distillation →
faithfulness judge-filter → SFT → GRPO.
GRPO setting
value
group size (G)
2
reward
format + valid-tool + schema-valid args + JSON-only + tool-match
KL reference
the SFT model
LoRA
attention q/k/v/o, r16/α32
group reward range
components sum to ≈[0, 1] per step
The GRPO LoRA is merged into the published GGUF weights (this repo).
Training data & provenance
Built from rdubwiley/agenda-parser-tool-traces: per-step {system, user, assistant} chat examples whose system message is the deployed agent's exact tool catalog + protocol. The source agenda packets are published in that dataset's source_packets/ folder; each trace row links to its source by meta.unit_id. Distilled from third-party teacher models (their terms may apply to generated text); source PDFs are public meeting records.
(lite = fast/small; medium = balanced; high = best quality. medium/high share the 26B-A4B base, fine-tuned independently and shipped at different quants.)
Evaluation
One agenda packet (oakland-1570) and a held-out task seed are excluded from training and reserved for a base-vs-fine-tuned A/B benchmark (objective protocol metrics — valid-JSON-action rate, clean-final_answer rate, tool-error rate — plus an LLM-judge of answer faithfulness, absolute and pairwise). See the project repo's sft/eval.py.
