alpha-chat-v11-m300-experimental

Alpha chat v11-m300 experimental

This is a versioned experimental checkpoint published for independent evaluation. It did not beat Alpha's V8 development reference and is not the selected production checkpoint.

V11 tested an all-token causal-language-modeling bridge over 10,862 reviewed synthetic conversations. The conversation candidates were generated with GPT-5.4 and independently reviewed with GPT-5.5. The model began from the V8 step-200 checkpoint; the architecture, tokenizer, chat template, and inference renderer were kept fixed. Symbiogenesis was disabled.

Measured result

On the unchanged 615-conversation development diagnostic, this checkpoint produced 615 nonempty, EOS-terminated responses, with 12 detected degenerate loops and a mean four-gram repeat rate of 0.0120652. In a reference-blinded GPT-5.5 review of 100 conversations, it received 15 PASS, 11 BORDERLINE, and 74 FAIL judgments. The V8 reference received 16 PASS, 9 BORDERLINE, and 75 FAIL judgments and ranked first overall. The reviewer selected no candidate as conversationally competent.

This artifact is therefore useful as a negative result: full-dialogue token supervision improved response initiation and stopping, but did not produce a reliable semantic gain and increased repetition relative to V8.

Exact identity

• Native checkpoint step: 300
• Native checkpoint SHA-256: 6226c1443741058089f110b89dfa341e0325851098d3aaf049a501c1ca3393f9
• Tokenizer SHA-256: c310343a185aecb572b8b6568b55179df248f4adec009d14a9496da354090b24
• Source commit used for training: 1dfdb813ecdaa72bfa51ec07757f4265372c03a7
• Blinded review SHA-256: 29355fb8a4e8093472b08f0bb4438964383749c00dd2be8faf625ea468a40a1a
• Quality gate: FAIL

Usage

from transformers import pipeline

chat = pipeline("text-generation", model="ajaxdavis/alpha-chat-v11-m300-experimental")
result = chat([{"role": "user", "content": "Hello. What are you thinking about?"}], max_new_tokens=64)
print(result[0]["generated_text"][-1]["content"])

The export uses standard LlamaForCausalLM files and requires no custom model code. Its exact chat template is included in chat_template.jinja. Do not interpret nonempty output or a low training loss as evidence of conversational competence.

The source, experiment contract, evaluation tools, and outcome record live in the Alpha2 repository.