FinnAI-Foundation-v3

FinnAI-Foundation-v3 is the third iteration of a verification-first LoRA fine-tune of LiquidAI/LFM2.5-VL-1.6B, prepared by FinnAI Foundation. It is the current best-measured candidate in the FinnAI-Foundation line.

What is in this repo

Path	What it is
FinnAI-Foundation-v3-merged-Q4_K_M.gguf	Q4_K_M GGUF for llama.cpp (697 MiB, 4.98 BPW)
mmproj-LFM2.5-VL-1.6b-Q8_0.gguf	Vision projector, Q8_0, required for multimodal llama.cpp
loss_history.json	Full training log (loss and eval loss per step)
MODEL_CARD_v3.md	This card

The LoRA adapter is published at Finn-Technologies/FinnAI-Foundation-LoRA-v3. Load it on LiquidAI/LFM2.5-VL-1.6B with peft and call merge_and_unload() to reproduce the merged fp16 checkpoint that produced this GGUF. The training data is published at Finn-Technologies/FinnAI-Foundation-100K-v3.

Method

Supervised distillation in the style of the MiMo-V2.6 recipe: teacher traces with concise <think> segments plus deterministic verifiers, distilled into LFM2.5-VL-1.6B with LoRA-SFT, then merged and quantized.

LoRA was applied at rank 32 / alpha 64 / dropout 0.05 over the attention, MLP, and convolution projection matrices (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj, fc1, fc2, linear, w1, w2, w3).

Training

• Base: LiquidAI/LFM2.5-VL-1.6B (SigLIP2 NaFlex 400M vision encoder, LFM2.5-1.2B backbone, 32k context)
• Hardware: single Kaggle T4, fp16, gradient checkpointing
• Steps: 3,590 (one epoch over 114,870 rows, batch 2, accumulation 16)
• Runtime: 33,146 s (9.2 h)
• Final train loss: 0.4739
• Eval loss: 0.37896 (step 1000) -> 0.37209 (2000) -> 0.36878 (3000) -> 0.36843 (3590)
• Checkpoints saved every 500 steps; final adapter saved at end of run

Data

FinnAI-Foundation-100K-v3: 114,870 training rows and a frozen 2,224-row evaluation split, 27.0% of rows carrying an image, 100% gold-verifier pass with 80.0% verifier coverage and zero answer-template leaks.

v3 is a targeted repair of the v2 iteration: it preserves every v2 row and the v2 general-domain upsample, then adds eight deterministic prompt variants for each of the 731 diagram-mermaid rows (731 -> 6,579 rows) to restore visual weight after v2's general-domain upsample.

Intended use

On-device macOS assistant work: OCR, document and table questions, diagram reading, code and cyber reasoning, and visual UI assistance. Not intended as a safety authority or a substitute for review in high-stakes decisions.

Evaluation

BENCHMARKS.md in this repo holds the current numbers. Scores are produced with llama.cpp Q4_K_M and the sampling parameters LiquidAI documents for the base model (temperature 0.1, min_p 0.15, repetition penalty 1.05, max_image_tokens=256). Rows that fail to run are counted as failures rather than dropped, and every scored row passes through a deterministic verifier.

License

Released under the LFM Open License v1.0, inherited from LiquidAI/LFM2.5-VL-1.6B. Review that license before commercial use.