luxopes/Cognix-1-Flash-2.6B-LoRA

Cognix 1 Flash

Created by LuxAI. Code-first SFT of the pinned LFM2.5-2.6B Base, not Instruct. This repository contains the LoRA adapter.

BF16 unquantized frozen base; rank 64, alpha 128. LoRA covers attention linears, convolution input/output projections and FFN linears. Tied embeddings/output head, norms and depthwise convolution remain frozen.

Processed 313,058,111 input tokens and 170,415,279 assistant-loss tokens. Full epoch: True. Stop reason: epoch. Maximum training length 32768; whole overlong examples removed AFTER LFM retokenization. No right truncation. Independent padded batch rows isolate attention AND convolution. Loss only on originally selected assistant turns, including their end-of-turn token. Code FIM examples are presented as chat missing-code tasks.

Use the included unchanged upstream chat_template.jinja and tokenizer. Reasoning uses ...; Pythonic tool calls use <|tool_call_start|>[function_name(argument='value')]<|tool_call_end|>. Tool results use plain content in the native tool role/template. Execute only allowlisted calls; never eval generated Python. No new vocabulary or identity special tokens were added. Identity examples teach Cognix 1 Flash, created by LuxAI.

Short acceptance checks, if completed, are included as separate JSON receipts. No inference benchmark scores or quality guarantees are claimed. Czech performance, tool reliability and long-context behavior still need evaluation after training. No claim that all 32k contexts will be practical on every 8 GB phone.

The included upstream LICENSE (LFM Open License v1.0) applies to the base/derivative; this model is NOT relabeled Apache-2.0. Preserve its terms and attribution. See BASE_MODEL_CARD.md, DATA_ATTRIBUTION.md, DATA_MANIFEST.json and conversion.json. HF visibility follows the owner's explicit authorization, if included; the original learning configuration is retained unchanged for checkpoint provenance. GGUF conversion is a separate CPU step after the GPU pod has stopped.

Chat identity template update

The native chat format now includes the identity Cognix 1 Flash, created by LuxAI and brief general guidance. Caller-provided system instructions are retained after this prefix. Special token IDs and all weight tensors are unchanged. See system_prompt.txt, chat_template.readable.jinja, and chat_template_update.json. Historical acceptance and LuxCore1 results predate this template change; no new score or quality improvement is claimed.