LoRA adapter for Qwen/Qwen3.5-4B, fine-tuned as the inference backbone for Mimir — a professional knowledge assistant focused on GRC (Governance, Risk, Compliance) and Information Security.

⚠️ Synthetic training data warning. The Round 2 curriculum in this repo (≈3% of the training mix) was generated by Claude Sonnet from hand-written seed patterns. These are not real user queries or verified professional content — they are synthetic training examples designed to teach specific tool-calling behaviours. Do not use this adapter for production decisions about actual regulatory, legal, or financial matters without expert review.

What this adapter does

It teaches Qwen3.5-4B three behaviours on top of its base chat ability:

Structured tool-calling — emits valid <tool_call>{...}</tool_call> blocks for a 20-tool Mimir toolset (mimir_search, mimir_thread, mimir_compare, mimir_ask, web_search, email_draft, jira_query, calendar_read, etc.), calling multiple tools in parallel where appropriate.
Governance compliance — respects "ask before write" rules: always calls mimir_ask before any file_create, mimir_save, email_draft(action=send), teams_message(priority=urgent), or jira_update.
Direct-answer discrimination — recognises queries that are settled professional knowledge (definitions, comparisons, conceptual explanations, professional norms) and answers directly, without any tool calls. This was the primary failure mode of v1 and the main target of v2.

Training data

Source	Examples	Role
Salesforce/xlam-function-calling-60k (gated, external)	58,716	General tool-calling patterns, reformatted into Qwen3 chat template
Round 2 curriculum (this repo → `data/round2_training.jsonl`)	630 raw, upweighted 3× to 1,890	Mimir-specific behaviours
└─ Should-Not-Call patterns	400	Direct-answer for knowledge questions
└─ Recency patterns	150	Parallel `web_search` + `mimir_search`
└─ Threading patterns	80	`mimir_thread`-first ordering
Combined training set	60,606	Round 2 share = 3.1%

The Round 2 data was generated by Claude Sonnet 4 from 46 hand-written seed patterns. Full generator script + seeds: Bharambe-NL/mimir-training.

Training config


Base model	`Qwen/Qwen3.5-4B` (bf16)
Method	LoRA (rank 128, alpha 256)
Trainable params	84.9 M (1.98 % of total)
Effective batch	64 (8 per-device × 8 grad-accum)
Optimizer	AdamW
LR schedule	2e-4 peak, cosine decay
Precision	bf16 + Flash-Attention 2
Sequence length	2048
Epochs	2
Hardware	1× NVIDIA H100 80GB HBM3
Wall time	4 h 00 m
Approx. cost	$15

Full training dynamics:

Training curves — loss, reduction %, grad norm, LR

Results

Training metrics (final, step 1894 / epoch 2)

Metric	Value
eval_loss	0.0663
eval mean_token_accuracy	98.10 %
train_loss	0.0659
train / eval gap	0.0004

Monotonic improvement across all 4 evals (step 500, 1000, 1500, 1894). No plateau, no overfit.

70-entry Mimir eval: v2 vs v1

Full eval set in eval/batch_*.md.

Metric	v1	v2	Δ
Family F (Should Not Call) pass rate	1 / 8 (12 %)	7 / 8 (88 %)	+76 pp
Governance compliance (of entries with tool calls)	68 / 69 (99 %)	57 / 57 (100 %)	+1 pp
UNPARSEABLE tool_call JSON	2 / 70	0 / 70	clean
Family F average tool calls	0.9	0.1	clean drop

The primary failure mode of v1 — over-eager tool-use for direct-answer questions — is largely fixed. One remaining Family F failure: MIMIR-075 (model still calls mimir_ask where it should answer directly about ISO 27001 certification timelines).

Tool density on the other five families shifted upward slightly (CVE 1.8→2.1, New Regulation 1.4→2.0, Product 1.3→1.9, TPRM 1.6→1.9, Threading 1.2→1.4), suggesting the model got more decisive on genuinely-tool-requiring queries rather than under-calling them.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base_model_id = "Qwen/Qwen3.5-4B"
adapter_id    = "Bharambe-NL/mimir-qwen3-4b-lora-v2"

tokenizer = AutoTokenizer.from_pretrained(base_model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

# Example inference
messages = [
    {"role": "system", "content": "You are Mimir ..."},  # full system prompt at link below
    {"role": "user",   "content": "What's the difference between ISO 27001 and SOC 2?"},
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,   # important: skips Qwen3's default reasoning scaffold
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.1,
    do_sample=True,
    pad_token_id=tokenizer.eos_token_id,
    eos_token_id=[
        tokenizer.eos_token_id,
        tokenizer.convert_tokens_to_ids("<|im_end|>"),  # stops cleanly after the turn
    ],
)
response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(response)

Full system prompt with 20 tool schemas: see SYSTEM_PROMPT in training/eval_runner.py.

Dependencies

Qwen3.5 is a hybrid architecture — linear-attention layers require these kernels to run on GPU:

pip install flash-linear-attention causal-conv1d

Without them, the model silently falls back to pure-torch CPU paths and inference is ~20 × slower. For pre-built wheels that match your torch version, see the causal-conv1d releases page.

Limitations

Synthetic training data (see warning at top). Do not rely on this model for production legal, regulatory, or financial decisions without expert review.
Slight regression on general xLAM eval vs v1 (+0.0004 eval_loss) — accepted trade for the 76-point Family F gain.
One remaining Should-Not-Call failure: MIMIR-075.
English only.
Trained on a narrow Mimir-specific tool surface (20 tools) — generalisation to other tool sets is unverified.
No adversarial/safety tuning beyond what is inherent to the base model.

Reproducibility

Complete training and eval infrastructure (scripts, seeds, eval set): Bharambe-NL/mimir-training.

Files in this repo:

adapter_model.safetensors, adapter_config.json — the LoRA adapter
tokenizer.json, tokenizer_config.json, chat_template.jinja — tokenizer artefacts from base model for convenience
data/round2_training.jsonl — the 630 synthetic training examples
eval/batch_A..F.md — full 70-entry Mimir evaluation set
training_curves.png — 4-panel dashboard of the training run

License

This adapter & code: Apache-2.0
Round 2 synthetic data (data/) & eval set (eval/): CC-BY-4.0
Base model (Qwen/Qwen3.5-4B): Qwen license (see base model card)
xLAM data (not redistributed here): see Salesforce/xlam-function-calling-60k license (CC-BY-NC-4.0, gated)

Citation

bibtex

@misc{mimir-qwen3-4b-lora-v2,
  author = {Bharambe, Sagar},
  title  = {Mimir-Qwen3-4B-LoRA v2: LoRA adapter for a GRC/InfoSec knowledge assistant},
  year   = {2026},
  url    = {https://huggingface.co/Bharambe-NL/mimir-qwen3-4b-lora-v2},
}