ℹ️ Use v2 instead. This v1 adapter is kept public for reference and before/after comparison. v2 fixes the primary failure mode — over-eager tool-use on direct-answer questions (Family F: 12 % → 88 % pass rate).

⚠️ Synthetic / external training data warning (applies to v2; v1 is xLAM-only). Use with expert review for any production legal/regulatory/financial decisions.

LoRA adapter for Qwen/Qwen3.5-4B, fine-tuned as a first-pass tool-calling model for Mimir (a GRC / InfoSec professional knowledge assistant).

Training

Base: Qwen/Qwen3.5-4B (bf16)
Method: LoRA (rank 128, alpha 256) — 84.9 M trainable params (1.98 %)
Data: Salesforce/xlam-function-calling-60k only — 58,716 examples reformatted to Qwen3 chat template
Hardware: 1× NVIDIA H100 80GB, bf16 + Flash-Attention 2
Config: 2 epochs, effective batch 64, LR 2e-4 cosine, seq 2048
Wall time: ~4 hours

Final training metrics (step 1836)

Metric	Value
eval_loss	0.0659
eval mean_token_accuracy	98.10 %
train_loss	~0.066

70-entry Mimir eval

Metric	v1
Family F (Should Not Call) pass rate	1 / 8 (12 %) — primary gap
Governance compliance	68 / 69 (99 %)
UNPARSEABLE tool_call JSON	2 / 70

v1 calls tools too aggressively on direct-answer questions. v2 addresses this specifically.

Full reproducibility + eval set: Bharambe-NL/mimir-training.

Usage

Same as v2 — just swap adapter_id = "Bharambe-NL/mimir-qwen3-4b-lora-v1". See the v2 model card for the full loading and generation snippet, including the critical enable_thinking=False and <|im_end|> stop-token settings.

License

Apache-2.0. Base model: Qwen license.