SelfLM 85M Chat 450K v1
This private research artifact contains the final deployable checkpoint from
the SelfLM compact-chat hackathon run. It uses custom PyTorch architecture code
from the SelfLM source repository and is not directly compatible with
transformers.AutoModelForCausalLM.
Architecture
- Approximately 85M parameters
- Decoder-only causal language model
- Standard grouped-query attention
- 2,048-token context window
- 16,384-token custom tokenizer
- Two MTP heads present; MTP was frozen during chat adaptation
- Final optimizer step: 4,745
Evaluation
- Held-out BPB128: 0.993996
- CORE16: 0.103576
- Product-chat gate: 4/10
- Status:
PASS_REVIEW_REQUIRED
The checkpoint can produce basic English but remains unreliable on factual and
arithmetic questions. Identity and tool execution must be implemented by the
deterministic host runtime. Do not use it for safety-critical, medical, legal,
financial, or autonomous tool decisions.
Files
checkpoint-final-step-4745.pt: weights-only deployable checkpointtokenizer.json: canonical SelfLM tokenizerselflm-config.json: exact architecture/runtime metadataevaluation-report.json: sanitized evaluation metrics
The exact-resume optimizer checkpoint and private generations are deliberately
excluded.
Usage
Clone the SelfLM source repository and use
scripts/serve_selflm_mini_chat_v1.py as documented in its README. The runtime
strictly validates checkpoint format selflm-compact-450k-full-ddp/1.0.0.
License and release status
No general public model license is granted. This repository is a private
hackathon judging/research artifact. Third-party data retains its original
licenses. Public or commercial deployment requires a separate review.
